functional MRIDiagnosis of depression based on resting state

Academic year 2018-2019

Master of Science in Biomedical Engineering

Master's dissertation submitted in order to obtain the academic degree of

Counsellor: Prof. dr. ir. Pieter van MierloSupervisors: Prof. dr. ir. Pieter van Mierlo, Prof. Chris Baeken

Student number: 01307496Gert Vanhollebeke

Acknowledgements

If a man has lost a leg or an eye, he knows he has lost a leg or an eye; but if he has losta self—himself—he cannot know it, because he is no longer there to know it.

Oliver Sacks

As working on and writing a master’s dissertation is no easy task, it is only normal to thank the peoplethat helped me.

First of all I would like to thank my promotors, Prof. dr. ir. Pieter van Mierlo and Prof. Chris Baeken.The flexibility and willingness they both showed, that made it possible for me to explore and workout my own idea as well as their guidance and help throughout the year made the whole dissertationprocess much more pleasing.

Next, I would like to thank my parents. Living together with a student, staying calm when thatstudent is annoying must be hard. My deepest thanks for the continuous support throughout the lastyear, and all years before this one. Special thanks to my father and sister for taking the time to readand correct this dissertation.

Thanks also to Debby Klooster and Toon Van de Maele for helping me with the structural featurespart of this dissertation.

Lastly I would like to thank my fellow students and friends who struggled together with me. Troubleshared is trouble halved, let’s have a beer when this all is over.

i

Permission for usage

The author gives permission to make this master dissertation available for consultation and to copyparts of this master dissertation for personal use. In all cases of other use, the copyright terms haveto be respected, in particular with regard to the obligation to state explicitly the source when quotingresults from this master dissertation.

Gert Vanhollebeke, Ghent, May 2019

iii

Diagnosis of depression based on resting statefunctional MRI

byGERT VANHOLLEBEKE

Master’s dissertation submitted in order to obtain the academic degree ofMaster of Science in Biomedical Engineering

Academic year 2018 - 2019

Supervisors: Prof. dr. ir. PIETER VAN MIERLO, Prof. CHRIS BAEKEN

Counsellor: Prof. dr. ir. PIETER VAN MIERLO

Department of Electronics and Information systemsChair: Prof. dr. ir. KOEN DE BOSSCHERE

Faculty of Engineering and ArchitectureGhent University

Abstract

While research uncovers new insights in the pathology of depression and de-fines new brain regions associated with the disease, the diagnosis of depressionstill remains a challenging task. Because of limited availability of psychologistsand psychiatrists there is a large diagnosis delay. This leaves some patients inneed of help, which can lead to the deterioration of the patient’s mental health.In this master’s dissertation, both structural MRI and resting state functionalMRI data from healthy controls and patients with depression are used to obtainfeatures in order to train a classifier capable of diagnosing depression. Threedifferent feature types - MRI volumetry, fMRI intensity and fMRI functionalconnectivity - are used as input for the classifier. Good results, over 90% accu-racy, were obtained using fMRI functional connectivity features and the com-bination of fMRI intensity and fMRI functional connectivity features resultedin accuracies of 94%, indicating that resting state functional MRI data can beused to reliably diagnose depression.

Keywords

Depression, Computer-aided diagnosis, MRI, Resting state fMRI, Machine learn-ing

1

1

Diagnosis of depression based on resting statefunctional MRI

Gert VanhollebekeSupervisors: Prof. dr. ir. Pieter van Mierlo, Prof. Chris Baeken

Abstract—While research uncovers new insights in the pathol-ogy of depression and defines new brain regions associatedwith the disease, the diagnosis of depression still remains achallenging task. Because of limited availability of psychologistsand psychiatrists there is a large diagnosis delay. This leavessome patients in need of help, which can lead to the deteriorationof the patient’s mental health. A computer-aided diagnosis toolusing both structural MRI and resting state functional MRI datais developed. Features from healthy controls and patients withdepression are used to obtain features in order to train a classifiercapable of diagnosing depression. Three different feature types -MRI volumetry, fMRI intensity and fMRI functional connectivity- are used as input for the classifier. Good results, over 90%accuracy, were obtained using fMRI functional connectivityfeatures and the combination of fMRI intensity and fMRIfunctional connectivity features resulted in accuracies of 94%,indicating that resting state functional MRI data can be used toreliably diagnose depression.

Index Terms—Depression, Computer-aided diagnosis, MRI,Resting state fMRI, Machine learning

I. INTRODUCTION

Depression is a common mental disorder resultingin a persistent saddened mood and anhedonia, possiblyaccompanied with other symptoms. Over 300 million peoplesuffer from depression worldwide, making it one of themost common mental illnesses [1]. Depression is not asingle disease, but a general name describing a multitude ofsymptoms. Many different causes have been defined.

The main method of diagnosing depression is a diagnosticinterview in which a professional psychologist or psychiatristexamines the patient to understand the symptoms the patientexperiences and asses the severity of the disease. Thisdiagnosis method is based on symptoms, which do notalways reflect the origin of the disease or any comorbiditiesthat are present. Many countries also do not have enoughmental health professionals, resulting in long waiting timesfor patients and inadequate care. Additional diagnosis toolscapable of diagnosing depression reliably and fast whilealso capable of diagnosing depression subtypes such asmedication- and treatment-resistant depression are needed.

Neuroimaging has proven to be useful in the diagnosisof neurological disorders such as epilepsy and multiplesclerosis, but is not yet used consistently in the diagnosisof mental disorders [2], [3]. Diagnosis tools for depressionand other mental illnesses based on neuroimaging techniquesthat achieve high accuracies have been developed, but the

features used for classification do not reflect aspects of brainanatomy and function that could be affected by depression,making the clinical validation of such diagnosis tools difficult.

In this paper, a computer-aided diagnosis tool based onanatomical MRI and resting state functional MRI (fMRI)scans from a data set of 106 people (60 healthy controls,46 depression patients) capable of diagnosing depression hasbeen developed. Section II describes the preprocessing steps,section III describes the different feature types and subtypesand defines the feature selection process.The features thathave been selected can easily be linked to different aspects ofbrain anatomy and function, increasing the clinical value ofthe diagnosis tool. Section IV describes the clinical relevanceof the found features, section V the classification trainingpipeline. Section VI describes the results and discussion,section VIII delineates the final conclusion that has beenreached.

II. FEATURE PREPROCESSING

All fMRI data is preprocessed using the CONN toolbox,which uses functions from the statistical parametric mappingsoftware toolbox [4], [5]. A preprocessing pipeline is selectedand adjusted to the needs of the data set.

All fMRI files are converted to the nifti file format, thefirst and last five scans are removed for signal equilibriumand signal dropout. Motion correction, slice timing correctionand coregistration are applied. Afterwards the time series arehigh pass filtered to remove scanner drift; this is done usingthe MATLAB toolbox from Anthony Barone [6]. Finally thedata is normalized to the MNI space and is smoothed usinga gaussian kernel with 6mm width.

III. FEATURE SELECTION

Three different feature types are investigated: intensityfeatures, connectivity features and structural features. Eachfeature type reflects a different aspect of the brain and couldpossibly show alterations in brain anatomy and function dueto depression. Each feature type contains multiple featuresubtypes.

The feature selection process consists of three parts: afeature calculation process, assessing group differences and afeature selection process. In the feature calculation processall possible features of each type and subtype are calculated

2

for each person in the data set separately. In the assessmentof group differences the feature values are averaged for eachgroup (healthy controls and depression group) separately andthe average depression group feature values are subtractedfrom the average healthy controls feature values. In thefeature selection process, the twenty features with the highestdifference between the average group feature values of bothgroups are selected as final features. The final features aretested using a two-pair t test to investigate their statisticalsignificance. Finally the features are normalized using thez-score.

A. Structural features

Two different subtypes of structural features are defined:cortical thickness and brain parcel volume. These features,that are obtained using the FreeSurfer software [9], reflectpossible anatomical changes that are present in people withdepression. Contrary to the intensity and connectivity features,no feature selection process is used as fewer possible featuresare available. All possible features are tested using a two-pair ttest for statistical significance. Only 19 left hemisphere corticalthickness, 11 right hemisphere cortical thickness and 6 parcelvolume features are statistically significant.

B. Intensity features

Intensity features reflect the average activity of the brainthrough time. These features are obtained from fMRI data andcould be compared (to a certain extent) to positron emissiontomography (PET) or single photon emission computedtomography (SPECT) imaging as they show metabolicprocesses such as glucose uptake within the body and brain.Two different feature subtypes are defined: absolute andrelative intensity features.

1) Absolute intensity features: Absolute intensity featuresshow the absolute activity in the brain and are calculated byaveraging all time series belonging to a certain brain parcel,resulting in an average activity of each brain parcel. Thesefeatures are prone to differences between patient scans asthey are not normalized. Global elevations or decreases inintensity between patient scans have a large influence on thefeature values.

2) Relative intensity features: Relative intensity featuresshow the relative activity in the brain and are calculatedby averaging all time series belonging to a brain parcel toa single value (similar to absolute intensity features). Theaverage activity value of each brain parcel is normalized usingformula 1 where Irel,j is the relative intensity value of brainparcel j, Iabs,j is the absolute intensity value of brain parcelj and Ibrain,avg is the average intensity value of the wholebrain, calculated by averaging all averaged time series in thebrain. This subtype is more resilient against global variationsof intensity between patient scans.

Irel,j =Iabs,j − Ibrain,avg

Ibrain,avg(1)

C. Functional connectivity features

Connectivity features reflect the possible changes in func-tional connectivity of the brain due to depression. Thesefeatures are obtained from the MRI data. Two connectivitymeasures are used: correlation and mutual information [7].These connectivity measures are chosen as they are undirectedand reflect functional connectivity in the time domain. Theinfluence of global signal regression on classification is alsoinvestigated. As global signal regression is subject to muchdiscussion the connectivity measures are calculated on botha global signal regressed data set and a non-regressed dataset [8]. No extensions, such as graph based features, areexplored from this feature type as the interpretability andclinical relevance of these types of features are low.

IV. CLINICAL RELEVANCE ASSESSMENT

The features are obtained using a data-driven approach,no knowledge about depression is used as a bias in thefeature selection process. The clinical relevance of the selectedfeatures is therefore not certain and is assessed. The structuralfeatures are first discussed. The intensity features will bediscussed as one group as most features from both subtypesare the same. Only the connectivity features calculated on theregressed data set will be discussed as they resulted in muchhigher classification accuracies.

A. Structural features

1) Cortical thickness: A total of 30 cortical thicknessfeatures are statistically significant. Nineteen features arelocated in the left hemisphere, eleven in the right hemisphere.In the people in the depression group, all cortical regionshave a decreased thickness, possibly pointing to neuronalatrophy due to depression. Several cortical regions thatare statistically significant are related to depression: theleft hemisphere rostral middle frontal gyrus, precentralgyrus, insula, precuneus, pars orbitalis, frontal pole, superiorfrontal gyrus, post central gyrus, caudal middle frontalgyrus, the right hemisphere frontal pole and superior frontalgyrus [15], [16].

2) Parcel volume: Six parcel volume features are statisti-cally significant. Both the left and right cerebellum hemisphereand left and right caudate nucleus are found, which are linkedto depression [17], [18].

B. Intensity features

Most of the features (absolute intensity features: 16 out of20 features, relative intensity features: 16 out of 20 features)lie within four brain regions: the left and right superior frontalgyrus and the left and right rostral middle frontal gyrus.All regions lie within the prefrontal cortex and are linkedto depression. All but two features show less activity in thedepression group when compared to the healthy controls,reflecting the hypoactivity commonly found in people withdepression [10].

3

C. Connectivity features

1) Correlation: Some features are possibly linked tobrain parcels afflicted by depression such as the connectionbetween the right anterior cingulate cortex and the leftorbitofrontal cortex and the connection between the leftand right prefrontal cortex [11]. Connections involvingthe precuneus could also be related to depression as theprecuneus is part of the default mode network: disturbances inthis network have been found in patients with depression [12].

2) Mutual information: Four brain regions are present inalmost all mutual information features: the right paracentralgyrus, the right inferior parietal gyrus, the right superior pari-etal gyrus and the brain stem. Both increased and decreasedfunctional connectivity have been found in the paracentralgyrus of people with depression, but its specific role in thedisease is not yet known [13]. Contrary to literature, which re-ports reduced functional connectivity in both the right inferiorand superior parietal gyrus, increased functional connectivityis found in this data set [14]. The brain stem contains severalnuclei that possibly are involved in depression, but the specificconnections found here (connections between the brain stemand the right inferior parietal gyrus, the right superior parietalgyrus and the right paracentral gyrus) have not been describedin literature.

V. CLASSIFIER TRAINING PIPELINE

To be able to compare the results of different classifiers, aclassifier training pipeline is defined and is shown in figure 1.A first selection of 46 people from the healthy controls is madeto avoid class imbalance when training a classifier. Secondlya train and test set is defined from the selected individuals.The 80-20 rule is used, resulting in a validation set of 19people and a training set of 73 people. The model is trainedand afterwards validated. Both the first selection and the train-test selection is performed using a random permutation. Thisensures that each time the pipeline is used, a different variationof the available data is used. Important to notice is that twovalidation sets are used to validate a trained classifier. Theofficial validation set, containing nineteen people from boththe healthy controls and the depression group, and an optionalvalidation set, containing fourteen people, all healthy controls.

VI. RESULTS

Two different categories of classifiers have been trained:single feature and combined feature classifiers. Single featureclassifiers are trained on a feature set containing a singlefeature subtype. Combined feature classifiers are trained on afeature set containing two or three subtypes belonging to onefeature type. Most results of the combined feature classifiersare lower than the single feature classifiers, the exception tothis is a classifier trained on de combined feature sets of theabsolute intensity, relative intensity, correlation with regresseddata and mutual information with regressed data features. Eachsingle feature classifier is trained sixty times in total. Threedifferent amounts of features (variable amount, 10 featuresand 20 features) are used as input for the classifier and each

feature amount is trained twenty times. This leads to a resultdistribution of twenty samples for each feature amount. Thefinal results are shown as the average accuracy, calculated fromthe best performing result distribution of the official validationset and the corresponding result distribution of the optionalvalidation set. The average standard deviation (SD) of thesevalidation sets is also shown. It is a measure for the variabilityof the result distributions: the higher the standard deviation,the lower the reliability of the feature set.

A. Single feature classifiers

Nine different single feature classifiers are trained usingthe nine different feature subtypes. The results are shown intable I. Several conclusions can be formed from the results.Firstly the results show that both the intensity features havea comparable accuracy and SD, showing that the absoluteintensity features did not suffer from any possible globalvariation of intensity between patients. Secondly the resultsshow that global signal regression has a significant positiveinfluence on the connectivity features. An increase in accuracyof ±6% for the correlation features and ±30% for the mutualinformation features is obtained when global signal regressionis used. Decreases in SD of both feature types also show theincreased quality of the features. Thirdly the results show thatthe structural features are not capable of accurately classifyingdepression. A reason for this could be the assumption thatall people in the depression group have suffered the sameform of depression, have taken the same medications andunderwent the same therapies, while this is in reality nottrue. Outliers were found in the intensity and connectivityclassifiers. A classifier with an accuracy of 90.9% is obtained(not shown in table I as this shows the mean accuracy) in boththe correlation and mutual information feature sets (calculatedwith the regressed data set).

TABLE IMEAN RESULTS OF THE SINGLE FEATURE CLASSIFIERS.

Feature type Acc. (%) Sens. Spec. PPV NPVLH thickness 59,6 ± 15,1 0.58 0.525 0.55 0.556RH thickness 56,3 ± 14 0.626 0.574 0.595 0.605Parcel volume 61 ± 20 0.558 0.504 0.535 0.528Abs. Int. 74,7 ± 10,5 0.762 0.718 0.72 0.75Rel. Int. 73,3 ± 11,3 0.761 0.714 0.73 0.744Corr. non regr. 77,5 ± 15,8 0.803 0.756 0.77 0.789Mut. Inf. non regr. 49,5 ± 19,9 0.485 0.431 0.46 0.456Corr. regr. 83,1 ± 7,3 0.858 0.807 0.815 0.85Mut. Inf. regr. 79,2 ± 7,5 0.768 0.817 0.805 0.78

B. Combined feature classifiers

Four different combined feature classifiers are trained: anintensity feature classifier, a connectivity feature classifier, astructural feature classifier and a intensity and connectivityclassifier. The intensity classifier is trained using a feature setthat contains both the absolute and relative intensity features.The connectivity feature classifier is trained using a featureset that contains both the correlation and mutual informationfeatures calculated from the regressed data set. The structuralfeature classifier is trained using all structural features. Theintensity and connectivity classifier is trained using the

4

Fig. 1. Visualization of the classification pipeline.

absolute intensity features, the relative intensity features,the correlation with regressed data features and the mutualinformation with regressed data features. These classifiers areonly trained twenty times with all features. The results areshown in table II. Multiple conclusions can be made from theresults. Firstly the results show that most combined featureclassifiers perform worse than the single feature classifiers.This is counter-intuitive as more features would normallyresult in a better classification. A possible explanation for thiscould be that each feature contains some unwanted noise andthat more features result in more unwanted noise, reducingthe classification accuracy. Secondly the results show that thereduction in classification accuracy, when compared to theresults of their respective subtypes, is much higher for theconnectivity feature classifier than for the other combinedclassifiers (the structural classifier even has a small increasein accuracy). This is again counter-intuitive as the single typeclassifiers using the same features have the highest accuracy.

The intensity and connectivity classifier has, contrary tothe other combined feature classifiers, a higher accuracy thanthe single feature classifiers. A classifier from this type isobtained with an accuracy of ±94.7% (not shown in table IIas this shows the mean accuracy). The increase in accuracycompared to both the single feature and other combinedfeature classifiers is explained by the fact that more featuresare used (80 features). The classifier is trained with the dataof 73 people (see section V), which is less than the amountof features used. SVMs are able to correctly train with morefeatures than samples, but are prone to overfitting. Furthervalidation is needed for this classifier type.

TABLE IIRESULTS OF THE COMBINED FEATURE CLASSIFIERS.

Feature type Acc. (%) Sens. Spec. PPV NPVStructural 61,5 ± 12,4 0.613 0.561 0.58 0.594Intensity 70,3 ± 14 0.738 0.684 0.695 0.728Connectivity 61,8 ± 13,7 0.688 0.629 0.645 0.672Int. and Conn. 88.7 ± 6.97 0.879 0.827 0.835 0.872

VII. DISCUSSION

The obtained results (best mean accuracy = ±88%, bestaccuracy = ±94.7%) are comparable or higher to those found

in literature [19], [20] [21]. The clinical relevance as wellas the easy interpretation of the used features, which is notfound in literature, makes the classifiers that are obtainedhighly relevant. Structural features can not yet be used forclassification, but the addition of information about the severityand duration of the depressive episodes of the patients couldsolve this problem. Both intensity and functional connectivityfeatures prove to be adequate for classification. While combin-ing feature subtypes from a single feature type did not lead toan increase in accuracy, the combination of both intensity andfunctional connectivity features resulted in the best performingclassifiers.

VIII. CONCLUSION

A computer-aided diagnosis tool has been developed usingboth anatomical and fMRI data to diagnose depression. A data-driven approach without any prior assumptions or knowledgeabout the known effects of depression has been used to selectfeatures for classification. The clinical relevance of the featureshas been assessed and all feature types contain features thatcan be linked to depression from a clinical point of view. Thisshows that, while using a different approach, the same brainregions that are influenced and change due to depression areidentified. High classification accuracies have been obtainedwhen the features are used to distinguish depression fromhealthy controls. The highest accuracies are obtained when theintensity and connectivity features are combined. As a generalconclusion, it can be stated that resting state fMRI data can beused to accurately predict depression and could be used in thefuture to aid mental health professionals for fast and reliablediagnoses thus reducing their workloads while also reducingwaiting times for patients.

REFERENCES

[1] World Health Organization. (2017). Depression and other common mentaldisorders: global health estimates (No. WHO/MSD/MER/2017.2). WorldHealth Organization.

[2] Olson, L. D., & Perry, M. S. (2013). Localization of epileptic foci usingmultimodality neuroimaging. International journal of neural systems,23(01), 1230001.

[3] Polman, C. H., Reingold, S. C., Banwell, B., Clanet, M., Cohen, J. A.,Filippi, M., ... & Lublin, F. D. (2011). Diagnostic criteria for multiplesclerosis: 2010 revisions to the McDonald criteria. Annals of neurology,69(2), 292-302.

5

[4] Whitfield-Gabrieli, S., & Nieto-Castanon, A. (2012). Conn: a functionalconnectivity toolbox for correlated and anticorrelated brain networks.Brain connectivity, 2(3), 125-141.

[5] Penny, W. D., Friston, K. J., Ashburner, J. T., Kiebel, S. J., & Nichols,T. E. (Eds.). (2011). Statistical parametric mapping: the analysis offunctional brain images. Elsevier.

[6] ”myfreqfilter”, Anthony Barone, The University of Texas at Austin,Institute for Geophysics.

[7] Cover, T. M., & Thomas, J. A. (2012). Elements of information theory.John Wiley & Sons.

[8] Saad, Z. S., Gotts, S. J., Murphy, K., Chen, G., Jo, H. J., Martin, A.,and Cox, R. W. (2012). Trouble at rest: how correlation patterns andgroup differences become distorted after global signal regression. Brainconnectivity, 2(1), 25-32.

[9] Fischl, B. (2012). FreeSurfer. Neuroimage, 62(2), 774-781.[10] George, M. S., Ketter, T. A., & Post, R. M. (1994). Prefrontal cortex

dysfunction in clinical depression. Depression, 2(2), 59-72.[11] Ramirez-Mahaluf, J. P., Perramon, J., Otal, B., Villoslada, P., & Compte,

A. (2018). Subgenual anterior cingulate cortex controls sadness-inducedmodulations of cognitive and emotional network hubs. Scientific reports,8(1), 8566.

[12] Utevsky, A. V., Smith, D. V., & Huettel, S. A. (2014). Precuneus is afunctional core of the default-mode network. Journal of Neuroscience,34(3), 932-940.

[13] Kenny, E. R., O’brien, J. T., Cousins, D. A., Richardson, J., Thomas,A. J., Firbank, M. J., & Blamire, A. M. (2010). Functional connectivityin late-life depression using resting-state functional magnetic resonanceimaging. The American Journal of Geriatric Psychiatry, 18(7), 643-651.

[14] Dutta, A., McKie, S., & Deakin, J. W. (2014). Resting state networks inmajor depressive disorder. Psychiatry Research: Neuroimaging, 224(3),139-151.

[15] Niu, M., Wang, Y., Jia, Y., Wang, J., Zhong, S., Lin, J., ... & Huang,R. (2017). Common and specific abnormalities in cortical thickness inpatients with major depressive and bipolar disorders. EBioMedicine, 16,162-171.

[16] Mackin, R. S., Tosun, D., Mueller, S. G., Lee, J. Y., Insel, P., Schuff,N., ... & Weiner, M. W. (2013). Patterns of reduced cortical thickness inlate-life depression and relationship to psychotherapeutic response. TheAmerican Journal of Geriatric Psychiatry, 21(8), 794-802.

[17] Baldacara, L., Borgio, J. G. F., Lacerda, A. L. T. D., & Jackowski,A. P. (2008). Cerebellum and psychiatric disorders. Brazilian Journal ofPsychiatry, 30(3), 281-289.

[18] Kim, M. J., Hamilton, J. P., & Gotlib, I. H. (2008). Reduced caudategray matter volume in women with major depressive disorder. PsychiatryResearch: Neuroimaging, 164(2), 114-122.

[19] J. R. Sato, J. Moll, S. Green, J. F. Deakin, C. E. Thomaz, and R.Zahn, “Machine learning algorithm accurately detects fmri signature ofvulnerability to major depression,” Psychiatry Research: Neuroimaging233, 289 (2015).

[20] M. J. Patel, A. Khalaf, and H. J. Aizenstein, “Studying depression usingimaging and machine learning methods,” NeuroImage: Clinical 10, 115(2016).

[21] M. Wei, J. Qin, R. Yan, H. Li, Z. Yao, and Q. Lu, “Identifying majordepressive disorder using hurst exponent of resting-state brain networks,”Psychiatry Research: Neuroimaging 214, 306 (2013).

Contents

Acknowledgements i

Permission for usage iii

Abstract iv

Extended abstract vii

Table Of Contents xii

List of Figures xix

List of Tables xxi

1 Introduction 1

1.1 Problem definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 MRI and fMRI 7

2.1 MRI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.1.1 Physical principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.1.1.1 Magnetic moment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.1.1.2 Total nuclear spin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2.1.1.3 The larmor frequency . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2.1.1.4 Relevance for MRI imaging . . . . . . . . . . . . . . . . . . . . . . . . 9

2.1.2 The MRI scanner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2.1.2.1 The homogeneous magnetic field . . . . . . . . . . . . . . . . . . . . 10

2.1.2.2 The rotating magnetic field . . . . . . . . . . . . . . . . . . . . . . . . 11

2.1.2.3 The gradient magnetic field . . . . . . . . . . . . . . . . . . . . . . . 12

2.1.3 MRI imaging sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.1.3.1 Echo planar imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.2 Functional MRI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2.2.1 The BOLD response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

xiii

CONTENTS

2.2.2 Advantages and limitations of fMRI . . . . . . . . . . . . . . . . . . . . . . . . 14

2.2.3 Resting state fMRI and task-related fMRI . . . . . . . . . . . . . . . . . . . . . 14

3 Machine learning 15

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

3.1.1 Supervised learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

3.2 Support vector machines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

4 Structural features 19

4.1 Available dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

4.2 Structural features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

4.2.1 FreeSurfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

4.2.1.1 Step 1: Normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

4.2.1.2 Step 2: Skull strip . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

4.2.1.3 Step 3: Subcortical segmentation . . . . . . . . . . . . . . . . . . . . 20

4.2.1.4 Step 4: Statistics calculation . . . . . . . . . . . . . . . . . . . . . . . 20

4.2.1.5 Step 5: White ma�er segmentation . . . . . . . . . . . . . . . . . . . 21

4.2.1.6 Step 6: Brain division . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

4.2.1.7 Step 7: Tesselation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

4.2.1.8 Step 8: Surface smoothing . . . . . . . . . . . . . . . . . . . . . . . . 21

4.2.1.9 Step 9: Inflation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

4.2.1.10 Step 10: Surface definitions . . . . . . . . . . . . . . . . . . . . . . . 21

4.2.1.11 Step 11: Spherical inflation . . . . . . . . . . . . . . . . . . . . . . . . 21

4.2.1.12 Step 12: Parcel labeling and statistics calculation . . . . . . . . . . . 21

4.2.2 Interpretation of the found features . . . . . . . . . . . . . . . . . . . . . . . . 22

4.2.2.1 Le� hemisphere thickness . . . . . . . . . . . . . . . . . . . . . . . . 22

4.2.2.2 Right hemisphere thickness . . . . . . . . . . . . . . . . . . . . . . . 23

4.2.2.3 Parcel volume . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

5 Functional features 25

5.1 Preprocessing process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

5.1.1 Step 0: File conversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

5.1.2 Step 1: Removal of the first and last scans for signal equilibrium . . . . . . . . 26

5.1.3 Step 2: Motion correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

5.1.4 Step 3: Slice timing correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

5.1.5 Step 4: Coregistration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

5.1.6 Step 5: High pass filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

5.1.7 Step 6: Normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

5.1.8 Step 7: Spatial smoothing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

5.2 Parcellation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

xiv

CONTENTS

5.2.1 Human brain atlas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

5.2.2 Human brain atlas resizing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

5.3 Intensity features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

5.3.1 Absolute intensity and relative intensity . . . . . . . . . . . . . . . . . . . . . . 30

5.3.2 Feature selection process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

5.3.2.1 Time averaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

5.3.2.2 Parcel averaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

5.3.2.3 Group averaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

5.3.2.4 Di�erence calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

5.3.2.5 Feature selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

5.3.3 Interpretation of the found features . . . . . . . . . . . . . . . . . . . . . . . . 33

5.3.3.1 Interpretation of the used tables . . . . . . . . . . . . . . . . . . . . . 33

5.3.3.2 Absolute intensity features . . . . . . . . . . . . . . . . . . . . . . . . 33

5.3.3.3 Relative intensity features . . . . . . . . . . . . . . . . . . . . . . . . 34

5.4 Connectivity features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

5.4.1 Functional connectivity measures . . . . . . . . . . . . . . . . . . . . . . . . . . 36

5.4.1.1 Selection process of connectivity measures . . . . . . . . . . . . . . . 37

5.4.2 Feature selection process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

5.4.2.1 Global signal regression . . . . . . . . . . . . . . . . . . . . . . . . . 38

5.4.2.2 Parcel simplification . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

5.4.2.3 Connectivity measure calculation . . . . . . . . . . . . . . . . . . . . 40

5.4.2.4 Group averaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

5.4.2.5 Di�erence calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

5.4.2.6 Feature selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

5.4.3 Interpretation of the found features . . . . . . . . . . . . . . . . . . . . . . . . 41

5.4.3.1 Features calculated with the non-regressed data set . . . . . . . . . . 41

5.4.3.2 Features calculated with the regressed data set . . . . . . . . . . . . 42

6 Classifier training 45

6.1 Classification training pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

6.1.1 Starting point . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

6.1.2 Class balancing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

6.1.3 Train-validation spli�ing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

6.1.4 Model training and validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

6.1.4.1 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

6.1.4.2 Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

xv

CONTENTS

7 Results 47

7.1 Data collection and presentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

7.1.1 Data collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

7.1.2 Presentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

7.1.2.1 Violin plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

7.2 Global results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

7.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

7.3.1 Single feature type classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

7.3.1.1 The le� hemisphere thickness classifier . . . . . . . . . . . . . . . . . 50

7.3.1.2 The right hemisphere thickness classifier . . . . . . . . . . . . . . . . 51

7.3.1.3 The parcel volume classifier . . . . . . . . . . . . . . . . . . . . . . . 52

7.3.1.4 The absolute intensity classifier . . . . . . . . . . . . . . . . . . . . . 54

7.3.1.5 The relative intensity classifier . . . . . . . . . . . . . . . . . . . . . . 55

7.3.1.6 The correlation with non-regressed data classifier . . . . . . . . . . . 56

7.3.1.7 The mutual information with non-regressed data classifier . . . . . . 58

7.3.1.8 The correlation with regressed data classifier . . . . . . . . . . . . . 59

7.3.1.9 The mutual information with regressed data classifier . . . . . . . . 61

7.3.2 Combined-feature classifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

7.3.2.1 The structural feature classifier . . . . . . . . . . . . . . . . . . . . . 62

7.3.2.2 The intensity feature classifier . . . . . . . . . . . . . . . . . . . . . . 63

7.3.2.3 The connectivity feature classifier . . . . . . . . . . . . . . . . . . . . 65

7.3.2.4 The intensity and connectivity feature classifier . . . . . . . . . . . . 66

8 Discussion 69

8.1 Part 1: Feature type specific . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

8.1.1 Intensity features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

8.1.2 Connectivity features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

8.1.3 Structural features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

8.2 Part 2: Feature type comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

8.3 Part 3: Combined feature classifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

8.3.1 The intensity feature classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

8.3.2 The connectivity feature classifier . . . . . . . . . . . . . . . . . . . . . . . . . . 71

8.3.3 The structural feature classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

8.3.4 The intensity and connectivity feature classifier . . . . . . . . . . . . . . . . . . 71

8.4 Part 4: Performance with respect to atlas level . . . . . . . . . . . . . . . . . . . . . . . 72

9 Conclusion 74

Bibliography 76

xvi

CONTENTS

Appendices 82

A MRI parameters of the fMRI data 83

xvii

List of Figures

1.1 Prevalence of depressive disorders in the world [1]. . . . . . . . . . . . . . . . . . . . . 2

1.2 Prevalence of depressive disorders with respect to age [1]. . . . . . . . . . . . . . . . . 2

2.1 Visualization of the angle (φ) between the magnetic moment vector (µ) of a nucleus(shown in red) and an external magnetic field (B) and the corresponding precession.10 8

2.2 Example of an MRI image. The di�erent tissues (grey ma�er, white ma�er, skin, air)can clearly be distinguished. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.3 Simplified representation of an MRI scanner [2]. . . . . . . . . . . . . . . . . . . . . . . 10

2.4 Random orientation of the hydrogen atoms without an external magnetic field (le�),net magnetization ML due to a magnetic field B0 (right). . . . . . . . . . . . . . . . . . 10

2.5 Visualization of the two components of the net magnetization M. . . . . . . . . . . . . 11

2.6 Visualization of the longitudinal (T1) and transverse (T2) recovery [3]. . . . . . . . . . 11

2.7 Visualization of the BOLD response with respect to time. The x-axis shows time, they-axis the change in oxy-/deoxyhemoglobin concentration in the blood in percentage. 14

3.1 A possible distribution of data points. Data points beloning to C1 are shown by reddots, C2 is shown by green dots. V1 and V2 denote two variables. . . . . . . . . . . . . 16

3.2 Visual presentation of the optimal hyperspace separating C1 and C2. The supportvectors are encircled. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

5.1 The complete preprocessing pipeline. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

5.2 Visual representation of the T1 relaxation of the tissue and the corresponding BOLDsignal. The vertical lines show the start of a new scan. (based on [4]) . . . . . . . . . . 26

5.3 Visualization of the initial presence of scanner dri� (a) and the result a�er high passfiltering (b). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

5.4 Visualization of the e�ect of normalization. . . . . . . . . . . . . . . . . . . . . . . . . . 29

5.5 Visualization of the e�ect of spatial smoothing. . . . . . . . . . . . . . . . . . . . . . . 29

5.6 The intensity feature selection process. . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

5.7 The connectivity feature selection process. . . . . . . . . . . . . . . . . . . . . . . . . . 38

6.1 The complete classification training pipeline. . . . . . . . . . . . . . . . . . . . . . . . . 46

7.1 Example of the di�erent possible results and violin plots. . . . . . . . . . . . . . . . . . 48

7.2 Violin plot of the global results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

xix

LIST OF FIGURES

7.3 Violin plot of the results of the le� hemisphere thickness classifier. . . . . . . . . . . . 50

7.4 Violin plot of the results of the right hemisphere thickness classifier. . . . . . . . . . . 52

7.5 Violin plot of the results of the parcel volume classifier. . . . . . . . . . . . . . . . . . . 53

7.6 Violin plot of the results of the absolute intensity classifier. . . . . . . . . . . . . . . . . 55

7.7 Violin plot of the results of the relative intensity classifier. . . . . . . . . . . . . . . . . 56

7.8 Violin plot of the results of the correlation with non-regressed data classifier. . . . . . 57

7.9 Violin plot of the results of the mutual information with non-regressed data classifier. 58

7.10 Violin plot of the results of the correlation with regressed data classifier. . . . . . . . . 60

7.11 Violin plot of the results of the mutual information with regressed data classifier. . . . 62

7.12 Violin plot of the results of the structural feature classifier. . . . . . . . . . . . . . . . . 63

7.13 Violin plot of the results of the intensity feature classifier. . . . . . . . . . . . . . . . . 64

7.14 Violin plot of the results of the connectivity feature classifier. . . . . . . . . . . . . . . 65

7.15 Violin plot of the results of the connectivity feature classifier. . . . . . . . . . . . . . . 66

8.1 Violin plot of the results of the connectivity feature classifier. . . . . . . . . . . . . . . 72

xx

List of Tables

4.1 Significant le� hemisphere thickness features. . . . . . . . . . . . . . . . . . . . . . . . 22

4.2 Significant right hemisphere thickness features. . . . . . . . . . . . . . . . . . . . . . . 23

4.3 Significant parcel volume features. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

5.1 Absolute intensity features (Atlas3). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

5.2 Relative intensity features (Atlas5). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

5.3 Functional connectivity measures [4]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

5.4 Correlation with non-regressed data set features (Atlas2). . . . . . . . . . . . . . . . . 41

5.5 Mutual information with non-regressed data set features (Atlas3). . . . . . . . . . . . . 42

5.6 Correlation with regressed data set features (Atlas3). . . . . . . . . . . . . . . . . . . . 43

5.7 Mutual information with regressed data set features (Atlas3). . . . . . . . . . . . . . . 44

7.1 Best results of the le� hemisphere thickness classifier. . . . . . . . . . . . . . . . . . . 50

7.2 Best results of the right hemisphere thickness classifier. . . . . . . . . . . . . . . . . . 51

7.3 Best results of the parcel volume classifier . . . . . . . . . . . . . . . . . . . . . . . . . 53

7.4 Best results of the absolute intensity feature classifier (Atlas3). . . . . . . . . . . . . . . 54

7.5 Best results of the relative intensity feature classifier (Atlas5). . . . . . . . . . . . . . . 55

7.6 Best results of the correlation with non-regressed data classifier (Atlas2). . . . . . . . . 57

7.7 Best results of the mutual information with non-regressed data classifier (Atlas3). . . . 59

7.8 Best results of the correlation with regressed data classifier (Atlas3). . . . . . . . . . . 60

7.9 Best results of the mutual information with regressed data classifier (Atlas3). . . . . . 61

7.10 Best results of the structural feature classifier . . . . . . . . . . . . . . . . . . . . . . . 62

7.11 Best results of the intensity feature classifier . . . . . . . . . . . . . . . . . . . . . . . . 64

7.12 Best results of the connectivity feature classifier . . . . . . . . . . . . . . . . . . . . . . 65

7.13 Best results of the intensity and connectivity feature classifier. . . . . . . . . . . . . . . 67

xxi

Chapter 1

Introduction

1.1 Problem definition

The World Health Organization (WHO) defines depression as "a common mental disorder, characterizedby persistent sadness and a loss of interest in activities that you normally enjoy, accompanied by aninability to carry out daily activities, for at least two weeks".1 It is estimated that more than 300 millionpeople worldwide su�er from some form of depression, making it the most prevalent mental disorder.2

Depression is not a single disease having a single origin, but rather a group of illnesses that result insimilar symptoms, varying both in duration and severity. The origin of depression is widespread, rang-ing from neurotransmi�er imbalances [5], medication side e�ects [6], substance use and abuse [7],childhood trauma [8] to mood changes due to chronic illness [9]. Recent studies report the majorinfluence of gut bacteria [10] and infection [11] as initiators of depression. Rapid changes in social,cultural and environmental aspects of the society are also blamed for the recent elevation in prevalenceand incidence of depression [12], [13].

Depression is mainly treated with psychotherapy and medication. More severe forms, mainly medication-resistant depression, can be treated using neurostimulation methods like transcranial magnetic stim-ulation (TMS), direct current stimulation (DCS) [14] or deep brain stimulation (DBS). DBS and TMSare relatively new types of treatment and their e�icacy as well as optimizations are still researched.Patients with very severe forms of depression that cannot (su�iciently) be treated with the previouslymentioned methods, can sometimes be helped with electroconvulsion therapy (ECT) [15].

The two major symptoms of depression (as defined by the Diagnostic and Statistical Manual of MentalDisorders, fi�h edition (DSM-5)) are a depressive mood and a lack of interest or pleasure in mostactivities that are present for at least two weeks [16]. It should be noted that it is not necessaryfor both symptoms to be present. Further symptoms such as weight change, irregular sleep pa�erns,motor abnormalities, increased feelings of guilt or worthlessness, decreased concentration and suicidalthoughts, ideations or a�empts are defined and could help the di�erential diagnosis and assess theseverity of the illness.

According to the WHO, depression is the leading cause of disability worldwide and a major contributorto the global burden of disease [1]. An estimated 5.1% of women and 3.6% of men, equaling around 322million people worldwide, su�er from some form of depression. The prevalence of depressive disordersthroughout the world varies and is shown in figure 1.1, the variation of depressive disorders prevalence

1. h�p://www.who.int/mental_health/management/depression/en/2. h�ps://www.who.int/news-room/fact-sheets/detail/depression

1

Problem definition

with respect to age is shown in figure 1.2.

In Belgium specifically, 700 000 people struggle with mental health issues yearly, a considerable amountof these people with depression [17], [18]. On average three people commit suicide in Belgium eachday, making it one of the countries with the highest suicide rates in Europe. Belgium has the 11th

highest suicide rate in the world.3 Only one in three people experiencing mental health issues reachout for professional mental help, showing the broader problems around mental health such as stigmaand lack of access to proper professional help.4

Figure 1.1: Prevalence of depressive disorders in the world [1].

Figure 1.2: Prevalence of depressive disorders with respect to age [1].

3. h�p://worldpopulationreview.com/countries/suicide-rate-by-country/4. h�ps://www.geestelijkgezondvlaanderen.be/feiten-cijfers

2

Introduction

Diagnosis of depression

The diagnosis of depression is done in a diagnostic interview such as the structural clinical interviewfor DSM (SCID) [19]. The duration of an interview varies based on the complexity of the disorder andthe ability of the patient to correctly describe the symptoms and takes on average between one andtwo hours.5

Diagnostic interviews are the main diagnosis tool for mental disorders such as depression and havehigh accuracy when performed by experienced psychologists and psychiatrists [20]. Self-administereddepression measures such as the Patient Health �estionnaire-9 (PHQ-9) and the Geriatric DepressionScale-15 (GDS-15) exist and could provide an useful first step for people experiencing mental health is-sues, but the accuracy of such tests is lower as people do not have the training and experience requiredto correctly recognize symptoms and severity and to distinguish between the di�erent disorders thatexist [21].

A diagnostic interview is limited by the fact that it is based on symptoms. The multiple origins aswell as the possible resistances to certain treatment options (medication-resistant and treatment-resistant depression) do not always translate to variations in symptoms and are thus not recognizablein a diagnostic interview. Because of this, first treatment of depression is focused on trail periods ofdi�erent medications until one is found that alleviates the symptoms adequately. This trail period canbe considerable and can even be futile when a patient has medication-resistant depression.

While the diagnosis of depression can be accurate [22], the waiting time for a diagnostic interviewcan be long. The VVP (Vlaamse vereniging voor psychiatrie) recently mentioned the fact that waitingtimes for an appointment with a mental health professional can be as much as 18 months [17].6 Aperson experiencing serious mental health problems is however o�en in urgent need of psychologicalor psychiatric guidance and a long waiting period may worsen the mental health of the person aswell as the possible outcomes. Limited access to psychological and psychiatric care has been linkedmultiple times to the rate of suicide.7

The limitations of the diagnostic interview as well as the long waiting times for professional helpgive rise to a need for additional diagnostic tools. Neuroimaging techniques are promisins as a newdiagnostic tool as they are routinely used to diagnose neurological diseases such as multiple sclerosis(MS) 8 and epilepsy 9 and could possibly be used to also diagnose depression. A major di�erencebetween the previously mentioned diseases and depression is that depression does not (yet) have awell defined origin within the brain (contrary to MS where brain lesions can be seen on a magneticresonance imaging (MRI) scan) and has no easily measurable neurological symptoms (contrary toepilepsy where seizures can be recognized in electroencephalography (EEG) as drastic changes inamplitude and frequency). This makes the use of neuroimaging to diagnose depression more relianton complex algorithms than on visual analysis of the physician.

Several imaging techniques, such as EEG and functional MRI (fMRI), have already been used in ana�empt to identify the di�erent brain structures related to depression and to define biomarkers10 fordepression diagnosis.

5. h�ps://www.verywellmind.com/structured-clinical-interview-25105326. h�ps://www.geestelijkgezondvlaanderen.be/feiten-cijfers7. h�ps://www.rand.org/research/gun-policy/analysis/supplementary/mental-health-access-and-suicide.html8. h�ps://www.mayoclinic.org/diseases-conditions/multiple-sclerosis/diagnosis-treatment/drc-203502749. h�ps://www.webmd.com/epilepsy/guide/electroencephalogram-eeg

10. h�ps://dictionary.cambridge.org/dictionary/english/biomarker: Biomarker: something, for example a gene or substance,that shows that a particular biological process or condition is present.

3

Objective

EEG has a high temporal resolution and can be used to examine the possible changes in frequencyand connectivity to develop a depression biomarker but has a low spatial resolution, making theidentification of influences from specific brain regions di�icult, if not impossible [23]. EEG basedbiomarkers that can diagnose people with depression with great accuracy have been developed; someresults report accuracies ranging from 85% up to 99% [24], [25], [26], [27]. The absence of a train-validation split of the data set or cross-validation from some of these papers as well as the small datasets (30 to 60 people) that are used by all papers limit the value of these biomarkers.

Depression biomarkers based on fMRI data have also been developed. Both resting state and task-related fMRI data are used to define biomarkers. Resting state biomarkers should reflect the changesin resting state brain networks and could point to changes in self referential thinking/rumination(negative self-image) due to depression while task-related fMRI biomarkers reflect reactions to stimulior tasks [28], [29]. Acuracies up to 92% have been reported for resting state fMRI based biomark-ers [30], [31], [32]. Task-related fMRI biomarkers have been reported with accuracies up to 95%[33],[34][35],[36].

Although resting state fMRI based depression biomarkers have a high accuracy, they all have twodisadvantages that limit their potential use in practice. The first limitation is that all biomarkers havebeen built using a limited data set. This lowers the possibility that the biomarker detects changes thatare due to depression and increases the chances of detecting imbalances in the data set instead. Itshould be noted that this is probably not the case, just that the possibility rises. A second limitationis that almost all biomarkers that have been developed are built using features that are not easilyinterpretable and thus provide li�le to no information about the underlying illness.

Both problems will be addressed in this master’s dissertation. A large data set of 46 patients withdepression and 60 healthy controls will be used to build a biomarker for depression. The features thatare explored will be easily interpretable for a physician.

1.2 Objective

The objective of this master’s dissertation is to build a computer-aided diagnosis tool capable ofdiagnosing depression based on an MRI scan of the brain, a resting state fMRI scan of the brainor a combination of both. The data set that is available contains 60 healthy controls and 46 patientsdiagnosed with depression. The master’s dissertation is composed of nine chapters and consists offive main parts: a theoretical introduction, structural feature selection, functional feature selecton,classifier training and result interpretation.

Theoretical introduction

In the first part the basic principles of MRI and fMRI will be explained as well as some principles ofmachine learning, specifically support vector machines.

Structural feature selection

The second part is the selection of the structural features. These features reflect the structural aspectsof the brain such as cortical thickness and volumes of di�erent brain regions. These features arecalculated from a T1 MRI scan using the FreeSurfer program [77].

Functional feature selection

The third part is the functional feature selection process. Two di�erent feature types are explored: in-tensity features and connectivity features. Intensity features are features that incorporate the generalactivity in the brain, connectivity features reflect the functional connectivity variability in the brain.

4

Introduction

A data-driven approach instead of a clinical approach is taken in the feature selection process. Noprior knowledge about the pathology of depression is used to influence the feature selection process.The final used features are selected only due to their significance in the data set. Clinical relevance ofthe features, if any, is assessed a�er the features have been selected.

Classifier training

The fourth part is classifier training. A classifier training pipeline is defined and the feature setsobtained in the third part are used as input for the training of a classifier. Due to the limited size ofthe data set support vector machines are the main machine learning algorithm that is used as classifier.

Firstly training is done using a single feature type as input. This makes it possible to assess theviability of each feature type as a distinguishing factor in the classification task. The final resultof this classification can give insight in possible underlying mechanisms of depression that may bepreviously not considered. General conclusions should be made very cautiously however, as the foundresults only reflect people in the used data set, not the general population.

At a later stage the most significant features of all feature types are combined and used to trainthe final classifier.

Result interpretation

The final part is the interpretation of the results found in part four. The accuracy of the di�erentcomputed classifiers is compared and results are discussed.

5

Chapter 2

MRI and fMRI

In this chapter, the neuroimaging techniques that are used to measure the brain activity are explained.In the first part of the chapter the basic principles of MRI will be clarified while the second part ofthe chapter will describe the underlying principles of fMRI. This chapter is mainly based on the booksMRI from picture to proton by DW McRobbie, EA Moore, MJ Graves and MR Prince [3] and Functionalmagnetic resonance imaging by SA Hue�el, AW Song and G McCarth [39].

2.1 MRI

Magnetic resonance imaging, also known as MRI, is an anatomical imaging technique widely used inmedicine to visualize the internal structures and organs of humans (and animals). The basic principleof MRI is nuclear magnetic resonance (NMR), which di�erentiates MRI from other medical imagingtechniques such as X-ray radiography, computed tomography (CT), positron emission tomography(PET), single photon emission computed tomography (SPECT) and ultrasound.

2.1.1 Physical principles

2.1.1.1 Magnetic moment

Every atomic nucleus is built up from two subatomic particles: neutrons and protons. Both particlesrotate around their own axis, creating a spin angular moment. This spin angular moment, S, is definedby a spin quantum number. As each particle can spin in only two possible directions, clock- andcounterclockwise, the spin quantum number has only two possible values: +1

2 and −12 .

Neutrons are electrically neutral, but protons do have a small positive electric charge. This electriccharge can be modeled as a small current, I. The circuital law of Ampère states that a moving electricalcurrent generates a magnetic dipole field B. This magnetic field is defined by its magnetic moment µ,given by formula 2.1.

µ = γ.S (2.1)

Here γ is the gyromagnetic ratio (HzT ) and S is the spin angular moment defined by the spin quantum

number. The magnetic moment can only occupy one of two possible states due to the limited possiblevalues of the spin angular moment.

7

MRI

2.1.1.2 Total nuclear spin

The total nuclear spin of a nucleus depends on the total amount of protons and neutrons from whichthe nucleus is built up. When the nucleus consists of an even number of both neutrons and protons,the total nuclear spin is zero. When a nucleus consists of an even amount of one subatomic particleand an uneven amount of the other subatomic particle, the total nuclear spin will have a half integervalue. When a nucleus consists of an odd amount of both subatomic particles the total nuclear spinwill have a full integer value.

When the total nuclear spin of a nucleus is di�erent from zero, the nucleus will have a magneticmoment. This magnetic moment is defined by formula 2.1, where the gyromagnetic ratio is nucleusspecific.11

2.1.1.3 The larmor frequency

When a nucleus with a magnetic moment di�erent from zero is objected to an external magneticfield (B), the nucleus will try to align with the magnetic field. Due to the fact that only two statesare possible, the alignment of the nucleus with regard to the external magnetic field is not perfect: asmall angle exists between the magnetic moment vector and the magnetic field vector.

When an angle exists between the magnetic field and the magnetic moment, the magnetic field exertsa torque on the magnetic moment of the nucleus, this torque (τ ) is defined by formula 2.2.

−→τ = −→µ x−→B = γ.

−→S x−→B (2.2)

Where −→τ is the torque vector, −→µ is the magnetic dipole moment,−→B the external magnetic field, γ

the gyromagnetic ratio and−→S the angular momentum vector. This principle is shown in figure 2.1.12

Because of the torque, the angular momentum vector will start to precess around the external mag-netic field vector. This precession occurs at a specific frequency, called the larmor frequency, given byformula 2.3.

ωL =1

S.sin(φ).γ.S.B.sin(φ) = γ.B (rad/sec) (2.3)

In this formula φ is the angle between B and µ (shown in figure 2.1).

Figure 2.1: Visualization of the angle (φ) between the magnetic moment vector (µ) of a nucleus(shown in red) and an external magnetic field (B) and the corresponding precession.10

11. h�p://nmrwiki.org/wiki/index.php?title=Gyromagnetic_ratio12. h�p://zerpoii.opentronix.com/?paged=4&tag=featured

8

MRI and fMRI

Formula 2.3 shows that the larmor frequency depends on both the strength of the external magneticfield and the gyromagnetic ratio of the nucleus. The gyromagnetic ratio is unique for every nucleustype, so every nucleus type has a unique larmor frequency.

2.1.1.4 Relevance for MRI imaging

A single hydrogen atom is built up from a single proton and an electron. As described in section 2.1.1.3,this proton will have a small magnetic moment and will thus precess when an external magnetic fieldis present.

Hydrogen atoms account for roughly 62% of all atoms present in the human body [40]. When thehuman body is subjected to an external magnetic field, all hydrogen atoms in the body will startto precess. The gyromagnetic ratio of hydrogen is 42.58MHz/Tesla.13 Note that not only hydrogenis subject to precession; other atoms such as carbon and nitrogen will also precess, but at anotherfrequency [41]. When the precessing hydrogen atoms are excited (see section 2.1.2.2), a signal willbe generated and can be measured. Di�erent tissues in the human body contain di�erent amountsof hydrogen atoms per volume. As each hydrogen atom generates a signal when excited, and morehydrogen atoms close together lead to a stronger signal, certain tissues generate stronger signalsthan others. Depending on the imaging sequence that is used to obtain the signal (see section 2.1.3),di�erent signal strengths will translate into di�erent grey values on the final MRI image. An exampleof an MRI image of a brain is given in figure 2.2.

Figure 2.2: Example of an MRI image. The di�erent tissues (grey ma�er, white ma�er, skin, air) canclearly be distinguished.

13. h�p://mriquestions.com/who-was-larmor.html

9

MRI

2.1.2 The MRI scanner

This section will explain the di�erent parts of an MRI scanner and the influence they have on thegeneration of an MRI image. A simplified version of an MRI scanner is shown in figure 2.3.

Figure 2.3: Simplified representation of an MRI scanner [2].

2.1.2.1 The homogeneous magnetic field

When no magnetic field is present, the magnetic moments of the hydrogen atoms in the body areorientated randomly, resulting in the absence of a net magnetization. The presence of a strong externalmagnetic field (B0) results in the alignment of the magnetic moments to the magnetic field. Twopossible alignments are possible: spin up and spin down. If a hydrogen atom has a spin up alignmentthe direction of the magnetic moment is parallel to the direction of the magnetic field, a spin downalignment means anti-parallel alignment. Because more hydrogen atoms are in a spin up state, a netmagnetization (M) whose direction is parallel to B0 is present. This principle is shown in figure 2.4.14

The magnet that is used in an MRI scanner to generate B0 is shown in figure 2.3 and is denoted as"Magnet". M is a static magnetic field and cannot be measured using a detection coil. Extra steps needto be taken in order to generate a measurable signal.

Figure 2.4: Random orientation of the hydrogen atoms without an external magnetic field (le�), netmagnetization ML due to a magnetic field B0 (right).

14. h�p://199.116.233.101/index.php/Physics_of_MRI

10

MRI and fMRI

2.1.2.2 The rotating magnetic field

The net magnetization M cannot be measured if it stays static. To change the net magnetization asecondary magnetic field is used. The secondary magnetic field is a rotating field constructed by acombination of two radio frequency magnetic fields perpendicular to each other with changing fieldstrength. The change in field strength follows a sine signal and both signals are o�set by 90 degrees.The magnets generating this magnetic field are shown as "radio frequency coil" in figure 2.3. If thefrequency of the rotating magnetic field is equal to the larmor frequency of a certain nucleus type,the net magnetization M of only that specific nucleus type will precess around B0.

Due to the precession the net magnetization can be subdivided in two components: a longitudinaland a transverse component. The longitudinal component of the net magnetization (ML) aligns withB0 while the transverse component (MT)is perpendicular to B0. This is shown in figure 2.5. The anglebetween the magnetization and its longitudinal component, denoted as θ in figure 2.5, is called theflip angle. The flip angle is dependant on the duration of the radio frequency burst generated by therotating magnetic field. A longer radio frequency burst results in a bigger flip angle.

Figure 2.5: Visualization of the two components of the net magnetization M.

The longitudinal part of the net magnetization will remain stationary and can still not be measured,but the transverse part of the magnetization is measurable as it precesses at the larmor frequencyand tries to align again with B0. This signal is measurable because it induces an electric current. Thepart of the scanner that measures this current is called "scanner" in figure 2.3. The realignment ofthe transverse magnetization with B0 is called relaxation. During the relaxation ML will recover untilit is completely restored (called longitudinal relaxation) and MT will decay until none is le� (calledtransverse relaxation). The relaxation of both magnetizations depends on di�erent mechanisms thatare independent of each other, resulting in di�erent relaxation times that are shown in figure 2.6. Therelaxation time of the longitudinal magnetization is called T1, the relaxation time of the transversemagnetization is called T2.

Figure 2.6: Visualization of the longitudinal (T1) and transverse (T2) recovery [3].

11

MRI

2.1.2.3 The gradient magnetic field

The homogeneous and rotating magnetic field makes it possible to generate a measurable signal,but all tissues in the body generate a signal at the same time and no comprehensible images can begenerated. To solve this problem gradient magnetic fields are used; they are generated by gradientcoils (see figure 2.3).

The gradient magnetic field generates a magnetic field that varies slightly in three dimensions. Thismagnetic field changes the local magnetic field strength of the homogeneous magnetic field slightly(see section 2.1.2.1). Each position of the body is now subjected to a slightly di�erent magnetic fieldstrength. As the larmor frequency depends on the gyromagnetic ratio of the nucleus, but also on theapplied magnetic field strength (see formula 2.3), the hydrogen atoms at each postion of the body nowhave a slightly di�erent larmor frequency. As nuclei only start to precess when the rotating magneticfield (see section 2.1.2.2) rotates at their larmor frequency, each position of the body can be targeteddirectly by changing the frequency of the rotating magnetic field. The position of the tissue thatgenerates a signal can be encoded in the signal it generates. Now images can be created.

It should be noted that this is a simplified and incomplete explanation of the MRI scanner. For furtherdetails the reader is referred to the book MRI from picture to proton [3].

2.1.3 MRI imaging sequences

Each tissue type has di�erent properties (such as amount of hydrogen atoms and the atoms in the lat-tice surrounding the hydrogen atoms) that influence the relaxation times. This di�erence in relaxationtime results in di�erences in the measured signals. This di�erence is used to generate images. Manydi�erent imaging sequences that are able to visualize di�erent aspects of the body by manipulatingthe T1 and T2 relaxations exist. These will not be explained here, but can be found in [3]. Only echoplanar imaging (EPI) will be mentioned shortly as it is the imaging sequence that is used to obtainfMRI data.

2.1.3.1 Echo planar imaging

Echo planar imaging is an imaging sequence that is able to capture a complete 2D slice using a singleradio frequency pulse generated by the rotating magnetic field. This is done by changing the gradientmagnetic fields (see section 2.1.2.3) while the generated signal from the tissue is measured. Thisreduces the scanning time needed to measure the complete brain from tens of minutes to seconds.The main drawback of this method is the reduced resolution that is obtained. An example of a highresolution MRI image is shown in figure 2.2, an example of an EPI image is shown in figure 5.4. Formore details the reader is referred to chapter 16, To BOLD go: new frontiers of MRI from picture toproton [3].

12

MRI and fMRI

2.2 Functional MRI

Functional magnetic resonance imaging, also known as fMRI, is a functional imaging technique.Functional imaging is a form of imaging that does not reveal the anatomy of organs and structures,but reveals processes active in a person’s body. The most common functional imaging techniques arefunctional MRI, positron emission tomography (PET), single photon emission computed tomography(SPECT), computed tomography perfusion imaging (CTPI) and near-infrared spectroscopy (NIRS).Functional imaging can provide information about illnesses that do not yet (or will never) alter theanatomy of a patient substantially enough that they could be detected using anatomical imaging. Anexample of this is the increased glucose uptake of small tumors in the body [42].

FMRI shows the activity of di�erent brain regions both spatially and temporally with a normal MRIscanner (see section 2.1.2) using the EPI imaging sequence (see section 2.1.3.1). The underlying prin-ciple of fMRI is explained in section 2.2.1, the advantages and limitations of fMRI are discussed insection 2.2.2.

2.2.1 The BOLD response

FMRI shows the activity of the brain both spatially and temporally. It does this by measuring thedi�erences in blood flow within the brain.

All cells in the body need energy to be able to perform functions. The creation of energy requiresboth oxygen and some energy source, mainly glucose. The citric acid cycle (also called Kreb cycle)will use the oxygen and energy source to produce adenosine triphosphate (ATP), the main energysource for the body. Neurons in the brain do not contain the necessary energy sources and oxygenthemselves, so if they are active these resources are brought to them via the blood vessels in the brainthrough a process called the hemodynamic response. Local brain activity results thus in local variationsof blood flow.

Oxygen is bound to haemoglobin in the blood and haemoglobin can exist in two possible states: oxy-genated and deoxygenated. Oxygenated haemoglobin (called oxyhemoglobin) has di�erent magneticproperties than deoxygenated haemoglobin (called deoxyhemoglobin) as oxyhemoglobin is diamag-netic and deoxyhemoglobin is paramagnetic. Blood traveling towards active neurons will contain ahigher percentage of oxyhemoglobin relative to blood that delivered the oxygen to the active neu-rons. At the location of the active neurons oxyhemoglobin will become deoxyhemoglobin as oxygenis delivered to the neurons. This change of oxyhemoglobin/deoxyhemoglobin concentration shi�sthe magnetic properties of the blood from more paramagnetic to more diamagnetic. The magneticproperty shi� of the blood will be higher in locations where neurons are active compared to locationswith inactive/less active neurons as more blood is delivered and more oxygen is given to the activeneurons. This di�erence in magnetic properties is called the blood-oxygen level dependant (BOLD)response and is measurable using an MRI scanner.

The BOLD response changes through time and follows the hemodynamic response function closely.Its course through time is shown in figure 2.7.15 Two observations can be made from this figure: thechange in the BOLD signal is small (2%) when a stimulus is applied and takes a long time (25 seconds)to recover completely. Another observation is that the peak in the BOLD signal does not align in timewith the stimulus but is delayed by around 8 seconds.

15. h�p://mriquestions.com/does-boldbrain-activity.html

13

Functional MRI

Figure 2.7: Visualization of the BOLD response with respect to time. The x-axis shows time, the y-axisthe change in oxy-/deoxyhemoglobin concentration in the blood in percentage.

2.2.2 Advantages and limitations of fMRI

The main advantage of fMRI when compared to other neuroimaging techniques is its high spatialresolution of around 1-2mm. The high spatial resolution makes fMRI a unique imaging techniquefor identifying functional brain regions. Caution however should be used with the use of fMRI. It isnot sure that the location where the oxygen is delivered in the brain corresponds perfectly with thelocation of the active neurons; the variations in the BOLD signal that are measured do not necessarilymatch perfectly with the associated brain regions. Another advantage of fMRI is its capability ofrecording brain activity of subcortical regions. Other (non-invasive) functional imaging techniquessuch as EEG and fNIRS are not capable of recording activity deep within the brain.

The major disadvantage of fMRI is its low temporal resolution when compared to other neuroimagingtechniques (0.5-1Hz). Scanning the whole brain using an MRI scanner can take several minutes whenan anatomical scan is needed16 and the EPI imaging sequence (see section 2.1.3.1) that is used to obtainfMRI data significantly improves the temporal resolution, at the cost of spatial resolution. Otherimaging techniques like EEG have a much higher temporal resolution (ms range). The low temporalresolution limits the research opportunities for fMRI. EEG and fMRI can however be recorded simul-taneously and the data can be combined resulting in both high temporal and spatial resolution [39].

2.2.3 Resting state fMRI and task-related fMRI

Two types of fMRI data are defined: resting state and task-related fMRI data. Restig state fMRI is thecapture of the BOLD signal changes when the person is not engaged in any activity. The person lieswith his eyes closed in the MRI scanner and does not think about anything special. The person is notallowed to fall asleep. Several brain networks have been defined using resting state fMRI data suchas the default mode network audio/visual networks and sensory/motor networks [39].

Task-related fMRI is the capture of the BOLD signal changes when a person performs certain tasks.The nature of these tasks is varied ranging from viewing images to counting tasks [35], [43]. This typeof fMRI data is used to investigate the progression of signals through the brain and identify brainregions related to di�erent tasks a person can perform.

16. h�ps://www.healthline.com/health/head-mri

14

Chapter 3

Machine learning

This chapter will explain some principles of machine learning and the machine learning techniquethat is used: support vector machines (SVMs). Not all aspects of machine learning will be explained,only the principles that are of importance in this master’s dissertation. This chapter is based on thebook Introduction to machine learning by Ethem Alpaydin [44].

3.1 Introduction

Machine learning is the scientific study dedicated to the development and optimization of modelsand techniques used by computers to perform specific tasks without any specific instructions. Thisapproach makes it possible to solve complex problems for which no obvious or simple instructionscan be defined. The models instead try to learn underlying pa�erns using available data sets or bytrial and error. Depending on the problem that needs to be solved, di�erent models and approachesare required. The learning technique that is used in this master’s dissertation is called supervisedlearning.

3.1.1 Supervised learning

This type of learning can be used when a mapping from input to output is needed and data thatcontains both the input and correct output, called labeled data, is available. The data set for thismaster’s dissertation is labeled data. The input data is the MRI and fMRI scans, the output is the classthey belong to: healthy controls or depression group. Many models can be trained with supervisedlearning such as linear regression, random forests, artificial neural networks and SVMs.

3.2 Support vector machines

An SVM is a machine learning technique that is a part of a group of machine learning models calledkernel machines. SVMs are used in this master’s dissertation because high classification accuraciescan be obtained with small data sets (tens of samples) while other machine learning techniques, suchas artificial neural networks, need much larger data sets (thousands of samples).

The general principle of SVMs will be explained by describing a simple classification problem. Twoclasses are present and two variables describe each data point belonging to one of the two classes.Class 1 will be denoted by C1, class 2 by C2. Data points belonging to C1 are labeled with -1, datapoints belonging to C2 with +1. A possible distribution of data points of both classes is shown infigure 3.1.

15

Support vector machines

Figure 3.1: A possible distribution of data points. Data points beloning to C1 are shown by red dots,C2 is shown by green dots. V1 and V2 denote two variables.

The samples represented in figure 3.1 can be wri�en as X = {xt,rt}. X is a 2*t matrix containing everydata point x coupled with its corresponding class label r, with t the amount of data points. The goal ofa classification problem is to predict the class label r of a data point given the variables x that describeit. A solution for this problem is to define a hyperplane which separates both classes. As the currentcase is defined in two dimensions, the hyperplane will be a line.

The hyperplane can be defined by the function described in formula 3.1. It presents the hyperplaneas a function of the available data points (x), each multiplied by some weight (w) where an o�set isadded (w0).

g(x) = wTx + w0 (3.1)

The hyperplane will be able to separate both classes when it obeys two constrains, shown in formula 3.2and 3.3.

wTxt + w0 ≥ +1 for rt = +1 (3.2)

wTxt + w0 ≤ −1 for rt = −1 (3.3)

These constrains demand that for every data point belonging to C1 the hyperplane will return a value≥ 1 while returning a value≤ -1 for every data point belonging to C2. This equation can be simplifiedto formula 3.4.

rt(wTxt + w0) ≥ +1 (3.4)

It should be noted that this is a tough constraint to obey. Not only do we want a correct separation(which would require rt(wTxt + w0) ≥ 0) but also that all points are some distance away from thehyperplane. The space between the defined hyperplane and the data points that are closest to it iscalled the margin. The best classification results will be obtained when the margin on both sides ismaximized. If this is not possible, so� margin SVMs need to be used; these will not be explained butcan be found in [44].

The maximization of the margins is obtained using formula 3.5 and describes the minimization of thenorm of the weight vector w. This minimization results in a description of the optimal hyperplane bythe least amount of data points necessary as only the data points closest to the hyperplane will have aweight greater than zero. This results in a description of the optimal hyperplane by only a small subsetof the initial present data points. These data points are called the support vectors. The optimization of

16

Machine learning

the weight vector is the "learning" part of this machine learning technique. A visual representation ofthe optimal hyperplane of the distribution shown in figure 3.1 is shown in figure 3.2. This example wasa 2D classification problem as two variables were given. This principle can be extended to N variables.

min

(1

2||w||2

)subject to rt(wTxt + w0) ≥ +1 (3.5)

Figure 3.2: Visual presentation of the optimal hyperspace separating C1 and C2. The support vectorsare encircled.

It should be noted that this explanation concerns the most simple case possible. More informationabout kernel models can be found in Introduction to machine learning [44].

17

Chapter 4

Structural features

This chapter describes the feature selection process that is used to obtain the structural features. Firstthe specifics of the data set that is used are discussed. Secondly the program FreeSurfer [77], usedto calculate the structural features, is described. Finally the clinical relevance of the found features isassessed.

4.1 Available dataset

The data used in this master’s dissertation is obtained from Prof. Chris Baeken, one of the promotors.

The complete data set consists of two groups: healthy controls and patients with depression. Thehealthy control group consists of 60 individuals, the depression patients group consists of 46 indi-viduals. The data set of the depressed patients consisted originally of fi�y right-handed patients, ofwhich 35 were female (average age 42 years, standard deviation (SD) = 12 years). All patients wereantidepressant free at the time and were at least stage I treatment resistant, meaning that all of themhad had at least one unsuccessful treatment trial with serotonin reuptake inhibitors or noradrenalinereuptake inhibitors. Further exclusion criteria were current or past history of epilepsy, neurosurgicalintervention, having a pacemaker, having a metal object in the brain, having undergone electrocon-vulsion therapy, alcohol dependence or suicide a�empts within 6 months before the start of study.Patients with co-morbidities such as bipolar disorder and psychosis were also excluded. Depressionwas diagnosed using the structured Mini-International Neuropsychiatric Interview. Four patientswere not used; one female patient due to a suicide a�empt (medication overdose), one female patientdue to spontaneous improvement of the condition, one male patient due to an extra neurostimulationsession and one patient (sex unknown) due to the absence of data. The healthy controls are matchedfor sex, age and education level. Further details regarding the data set can be found in [45].

The healty control group will be called healthy controls, the patients with depression the depressiongroup.

Every person in the whole data set (106 people in total) has two di�erent MRI scans: a T1 weightedgradient echo (GE) scan using the MPRAGE protocol [46] and an fMRI scan using the EPI imagingsequence.

19

Structural features

4.2 Structural features

Structural features reflect possible anatomical changes due to depression and are obtained from theanatomical T1 images of the data set. The program that is used to obtain these features is calledFreeSurfer [77].

Two feature subtypes will be explored: cortical thickness and parcel volume. Cortical thickness refersto the thickness of the gray ma�er (in mm) of certain parcels. Parcel volume refers to the total volume(in mm3) of certain parcels. FreeSurfer uses a di�erent brain atlas than the Lausanne brain atlas, sothe structural features will not reflect properties of the same regions as the intensity and connectivityfeatures. Less di�erent regions are defined than in the Lausanne brain atlas: 35 cortical thicknessparcels are defined for each hemisphere and 39 parcel volumes are defined. Due to the low amount ofpossible features compared to the other two feature types, no feature selection process will be used.Instead the significance of each possible feature will be tested by a two-pair t test. Any significantfeature will be used in a feature set. Three feature sets are defined: a le� cortical thickness feature set,a right cortical feature set and a parcel volume feature set. The clinical relevance of these features isdiscussed in section 4.2.2, the results of the feature sets are shown in section 7.3.1 and are discussedin section ??. First the FreeSurfer program will be clarified.

4.2.1 FreeSurfer

A specific function from FreeSurfer is used, called recon-all. As it consists of a total of 29 steps17, thesteps will be clarified in a simple way and some steps will be put together as they are part of a largerstep.

4.2.1.1 Step 1: Normalization

The first step in the FreeSurfer workflow is normalization. The anatomical MRI scan will be trans-formed into the MNI305 atlas space using an a�ine transformation. Intensity correction is also applied.A second intensity correction will be performed a�er step 4 as the exclusion of the skull improvesintensity correction.

4.2.1.2 Step 2: Skull strip

The second step is called skull strip. As its name implies, the skull will be removed from the scan,leaving only the brain and neck.

4.2.1.3 Step 3: Subcortical segmentation

The third step is defining and segmenting the subcortical regions. This is done in multiple steps wherethe neck is stripped and several registrations to templates are made. This step ends with segmentedand labeled subcortical regions.

4.2.1.4 Step 4: Statistics calculation

The fourth step is the calculation of the statistics of the subcortical parcels. The volume of thesesubcortical parcels, used as parcel volume features, are calculated in this step.

17. h�ps://surfer.nmr.mgh.harvard.edu/fswiki/recon-all

20

Structural features

4.2.1.5 Step 5: White ma�er segmentation

The fith step is the segmentation of the white ma�er. The constraints used for this segmentationare the intensity di�erence between gray and white ma�er, the immediate neighboring voxels aroundeach voxel and the inherent smoothness of the border between white and grey ma�er.

4.2.1.6 Step 6: Brain division

The sixth step is the division of the brain in the le� and right hemisphere, the cerebellum and themidbrain.

4.2.1.7 Step 7: Tesselation

The seventh step is the tesselation of both hemispheres. The surface of each hemisphere is approxi-mated using a finite element method where the border, defined in step 5, is approximated using smalltriangles. The smallest edges of the triangles have the same length as the side of a voxel. A surface isdefined this way.

4.2.1.8 Step 8: Surface smoothing

The eight step is the smoothing of the surface defined in step 7. As the surface follows the voxels faceswhich define the surface the edges will be perpendicular to each other. Smoothing the edges reducesthe angle and makes the surface smoother.

4.2.1.9 Step 9: Inflation

The ninth step is the inflation of the smoothed surface. The surface will be inflated to smoothen thegyri and sulci and the transformation for each vertex and edge is calculated. When the inflated surfaceis obtained, it will be checked for any defects that are present due to errors in previous steps.

4.2.1.10 Step 10: Surface definitions

The cortical thicknesses are defined by aligning the inflated surface with the grey-white ma�er borderpresent in the original anatomical T1 MRI scan. A second surface that defines the pial surface iscreated.

4.2.1.11 Step 11: Spherical inflation

The inflated surface is further inflated until it becomes a sphere. This spherical surface is matchedto a spherical atlas defining the di�erent brain regions. Alignment is performed based on matchingfolding pa�erns of the brain and a�erwards using small scale pa�erns.

4.2.1.12 Step 12: Parcel labeling and statistics calculation

The di�erent parcels will be labeled and the statistics of of the parcel will be calculated. The corticalthicknesses of these parcels, used ase le� and right hemisphere thickness features, are calculated inthis step.

21

Structural features

4.2.2 Interpretation of the found features

A total of three feature sets (le� hemisphere thickness, right hemisphere thickness and parcel volume)have been defined and will be discussed. The results of the classification process can be found insection 7.3.1.

Interpretation of the used tables

The number column defines the significance of the feature (here the features with the lowest p-valuewhen a double-pair t test is performed). The feature with number 1 has the lowest p-value of allfeatures. The feature with number 10 had the tenth lowest p-value of all features. The parcel columndefines the name of the parcel whose thickness or volume is linked with the feature. The sign columndefines which group (healthy controls or depression group) has the biggest cortical thickness or parcelvolume. A "+" sign denotes that the healthy controls have on average a bigger cortical thickness orparcel volume, a "-" sign denotes that the depression group has on average a bigger cortical thicknessor parcel volume. It should be noted that this does not occur.

4.2.2.1 Le� hemisphere thickness

Only nineteen features are found to be statistically significant, they are shown in table 4.1. Reducedcortical thickness of multiple brain regions defined by the features have been reported in literature.The brain regions from which a reduced cortical thickness is closely linked with depression (as indi-cated by prof. Baeken) are the rostral middle frontal gyrus (feature 2), the precentral gyrus (feature6), the insula (feature 7), the precuneus (feature 10), the pars orbitalis (feature 14), the frontal pole(feature 16), the superior frontal gyrus (feature 17), the post central gyrus (feature 18) and the caudalmiddle frontal gyrus (feature 19) [78], [79], [80]. The presence of several brain regions associated withdepression shows the clinical relevance of the le� hemisphere thickness features. The results of theclassifiers trained with the le� hemisphere features are shown in section 7.3.1.1 and are discussed insection ??.

Table 4.1: Significant le� hemisphere thickness features.

Number Parcel Sign1 Pars opercularis +2 Rostral middle frontal gyrus +3 Superior temporal gyrus +4 Mean le� hemisphere thickness +5 Supramarginal gyrus +6 Precentral gyrus +7 Insula +8 Pars triangularis +9 Inferior temporal gyrus +10 Precuneus +11 Inferior parietal gyrus +12 Middle temporal gyrus +13 Lateral orbitofrontal cortex +14 Pars orbitalis +15 Fusiform +16 Frontal pole +17 Superior frontal gyrus +18 Post central gyrus +19 Caudal middle frontal gyrus +

22

Structural features

4.2.2.2 Right hemisphere thickness

Only eleven features are found to be statistically significant, they are shown in table 4.2. Only twoof them can consistently be linked with depression (as indicated by prof. Baeken): the frontal pole(feature 5) and the superior frontal gyrus (feature 10) [79], [80]. From a clinical relevance point ofview, the right hemisphere features are thus less significant than those of the le� hemisphere. Thisobservation, together with the fact that less statistically significant features have been found in theright hemisphere could show that a di�erence exists between the le� and right hemisphere. Animbalance between a hypoactive le� and a hyperactive right hemisphere in people with depression hasbeen reported multiple times and these observations could also point towards this imbalance [52], [53].

Table 4.2: Significant right hemisphere thickness features.

Number Parcel Sign1 Supramarginal gyrus +2 Inferior parietal gyrus +3 Pars triangularis +4 Mean right hemisphere thickness +5 Frontal pole +6 Inferior temporal gyrus +7 Superior temporal gyrus +8 Superior parietal gyrus +9 Pars opecularis +10 Superior frontal gyrus +11 Middle temporal gyrus +

4.2.2.3 Parcel volume

Only six features were found to be statistically significant, they are shown in table 4.3. The first twofeatures are both hemispheres of the cerebellum, indicating that the biggest change in volume betweenthe healthy controls and the depression group could be a decrease in volume in the cerebellum forpeople with depression. Reduced cerebellar volume and cerebellar atrophy in people with depression isdescribed in literature and an involvement of the cerebellum in several psychiatric disorders includingdepression is suspected and investigated [81], [82], [83]. The third and fourth features are both partsof the caudate nucleus. The involvement of the caudate nucleus in depression has been proposed assome diseases involving the caudate nucleus, such as caudate infarcts or Huntington’s disease, giverise to depressive symptoms. Reduced volume of both caudate nuclei has been reported in people withdepression [84], [85]. Reduced caudate volume has also been reported in other psychiatric disorderssuch as obssesive compulsive disorder [86]. Reduced thalamic volume is also linked with depres-sion [87]. From a clinical relevance point of view, the parcel volume features are highly significant.The results of the classifiers trained with the parcel volume features are shown in section 7.3.1.3 anddiscussed in section ??. The sixth feature is closely linked with the first two features.

Table 4.3: Significant parcel volume features.

Number Parcel Sign1 Right cerebellum cortex +2 Le� cerebellum cortex +3 Right caudate nucleus +4 Le� caudate nucleus +5 Right thalamus +6 Right cerebellum white ma�er +

23

Chapter 5

Functional features

This chapter will describe the calculation and clinical validation of the functional features. Twodi�erent functional features are calculated: intensity and connectivity features. Intensity featuresreflect the activity of the brain, connectivity features reflect the functional connections within thebrain. Firstly the preprocessing process will be described, secondly the feature selection process isexplained, thirdly the clinical validation of the obtained functional features is discussed.

5.1 Preprocessing process

This section will describe all preprocessing steps taken to prepare the data set for feature selection.The process is shown in figure 5.1. Several preprocessing steps can introduce errors in the data. Tocounter this, visual quality checks are performed. This is done by viewing the produced results of eachpreprocessing step of around ten people from each group.

Figure 5.1: The complete preprocessing pipeline.

25

Preprocessing process

5.1.1 Step 0: File conversion

The MRI scanner used to obtain the scans is a Siemens Magnetom TrioTim syngo MR B17 scanner.The file format of the scanner is the digital imaging and communications in medicine (DICOM) withfile extension .IMA. The zeroth preprocessing step is the conversion from the Siemens DICOM formatto the neuroimaging informatics technology initiative (NIfTI) format. This step is done by the SPMtoolbox ([37]) for further processing as the toolbox uses the ni�i format. This step is performed usingthe import.dicom function from the SPM toolbox.18

5.1.2 Step 1: Removal of the first and last scans for signal equilibrium

The first five scans from each patient are dismissed, this because the BOLD signal (see section 2.2.1)has not yet been stabilized due to an incomplete T1 relaxation (see section 2.1.2.2 and figure 2.6). Thesescans are called dummy scans. This principle is shown in figure 5.2.

The last five scans are also dismissed, this because the final scans of some patients show a significantdrop in voxel value throughout the whole scan. The origin of this signal drop is unknown.

Figure 5.2: Visual representation of the T1 relaxation of the tissue and the corresponding BOLDsignal. The vertical lines show the start of a new scan. (based on [4])

5.1.3 Step 2: Motion correction

Even though people are instructed to lie still when a scan is taken, small movements are inevitable asmovements due to breathing and cardiac pulse also influence the location of the patient within thescanner. To correct these small movements, motion correction is applied.

Motion correction is done using a six parameter rigid body transformation. The first three parametersdescribe the rotation within the 3D space, the last three the translation. The appropriate rotation andtranslation values, which are unique for each scan, are calculated using a least squares approach [47].This preprocessing step is performed using the SPM realign function.19 This preprocessing step andall folowing steps aside from the high pass filtering (see section 5.1.6), are performed using the CONNtoolbox [38].

18. h�ps://en.wikibooks.org/wiki/SPM/Importing_data_from_the_scanner19. h�ps://en.wikibooks.org/wiki/Neuroimaging_Data_Processing/Realignment

26

Functional features

5.1.4 Step 3: Slice timing correction

The next step in the preprocessing process is slice timing correction. This step is needed becauseit is assumed that a single EPI (see section 2.1.3.1) scan of the head is done instantaneously. Inpractice however, this is not correct. The EPI sequence used to obtain the fMRI data, scans the headin a sequence of several 2D slices. The acquisition of the 2D slices is not simultaneous but is donesequentially. This results in a single 3D scan where the values of each slice are obtained at a slightlydi�erent time.20

The solution to this problem is the interpolation of the measured values to correct for the time dif-ferences between the acquisitions, called slice timing correction. The time di�erences are defined bythe repitition time (TR) and acquisition time (TA) parameters of the EPI sequence. The TR parameterdefines the time between radio frequency bursts generated by the rotating magnetic field (see sec-tion 2.1.2.2). The TA parameter defines the time di�erence between the start of the first and last sliceacquisition and can be calculated by formula 5.1.21 The values for TR can be found in appendix A.

TA = TR− (TR

nSlices) (5.1)

The last two parameters that are needed are the reference slice and the slice order. The referenceslice is the slice that is firstly scanned and is dependent on the slice order ie. the order in whichthe di�erent slices are acquired. The used slice order is an interleaved slice order (see "Series" inappendix A), meaning that first all even or uneven slices are acquired a�er which the other half arecaptured. As the scanner used to obtain the fMRI data is a Siemens Magnetom TrioTim syngo MR B17(see appendix A), the slice order is dependent on the amount of slices. If the total number of slices iseven, the even slices are captured first; if the total number of slices is uneven, the uneven slices arecaptured first.22 The amount of slices is 40 (for this data set, see appendix A) so the even slices willbe captured first. This also defines the reference slice as slice two. This part of the preprocessing partis done using the temporal.st function from SPM, built into the preprocessing pipeline of the CONNtoolbox [38].

5.1.5 Step 4: Coregistration

The next step in the preprocessing is coregistration. Coregistration is the alignment of the functionaldata (fMRI data) to the structural data (or vice versa) so that they share the same coordinate space.This means that the time series in the fMRI data are now linked spatially with their correspondingbrain region in the structural MRI. Coregistration is a patient-specific preprocessing step where thefunctional data of each patient is mapped to their personal T1 MRI scan. Coregistration is performedusing a six parameter rigid body transformation (similar to motion correction, see section 5.1.3) wherethree parameters describe the applied translation and the other three the applied rotation. Thecoregister fuction from the SPM toolbox is used, which is built into the CONN toolbox [38].

5.1.6 Step 5: High pass filtering

The next step is high pass filtering and is done to eliminate scanner dri�. Scanner dri� is the intro-duction of a low frequency signal (0 - 0.01Hz) into the time series captured during an MRI scan. Theorigin of scanner dri� is a change in resonant frequency (see section 2.1.1.3) of the hydrogen protons

20. h�ps://en.wikibooks.org/wiki/SPM/Slice_Timing21. h�ps://andysbrainblog.blogspot.com/2012/11/slice-timing-correction-in-spm.html22. h�ps://www.siemens-healthineers.com/magnetic-resonance-imaging/magnetom-world/clinical-corner/application-

tips/slice-order-fmri.html

27

Preprocessing process

which is induced by changes in field strength of the static magnetic field (see section 2.1.2.1) [39].Even though scanner dri� is the main source of low frequency noise, other sources also influence theacquired signal [48].

The filter that is implemented has a bandstop frequency of 1128Hz and a bandpass frequency of 1

120Hz.The presence of scanner dri� and the result of the filtering is shown in figure 5.3. The functionmyfreqfilter23 is used for the implementation of the high pass filter.

(a) Original time series. (b) Filtered time series.

Figure 5.3: Visualization of the initial presence of scanner dri� (a) and the result a�er high passfiltering (b).

5.1.7 Step 6: Normalization

Normalization is the second registration step (a�er coregistration) in the preprocessing process. Againa body transform will be applied to the functional data of each person so that all people in the data setshare a common coordinate space. This is necessary to allow comparisons between patients and findcommon characteristics on a group level. The coordinate space to which the data will be registered isthe Montreal Neurological Institiute (MNI) space [49]. This step is performed by wrapping the personalT1 MRI scan of each patient to a the MNI template scan. Secondly the calculated transformation fromthis wrapping is applied to the functional data of each person.

A first step in the normalization process is segmentation, where personal probability maps for air,skull, cerebrospinal fluid (CSF), white ma�er and gray ma�er are made. These probability maps arecalculated from the T1 MRI scans of each person.

The second step is linear registration. The personal probability maps are aligned to template prob-ability maps using an a�ine transformation, the template probability maps are located in the MNIcoordinate space. The a�ine transformation has twelve parameters: three rotation parameters, threetranslation parameters, three zoom parameters and three shear parameters. The rotation and trans-lation parameters align the personal probability maps to the template probability maps; the zoomand shear parameters alter the shape of the personal probability maps so that they are as similar aspossible to the template probability maps.

The third and final step of the normalization process is non-linear registration. This registration isnecessary as not all person-specific variations, such as unique folding of gyri and sulci, can be correctedusing a�ine transformations. The non-linear registration is performed using a linear combination of

23. "myfreqfilter", Anthony Barone, The University of Texas at Austin, Institute for Geophysics.

28

Functional features

discrete cosine basis functions. The result of this step is shown in figure 5.4. This step is performedusing the segment and normalize function from SPM, which is used by the CONN toolbox [38], [37].

(a) Functional data before normalization. (b) Functional data a�er normalization.

Figure 5.4: Visualization of the e�ect of normalization.

5.1.8 Step 7: Spatial smoothing

The final step in the preprocessing process is spatial filtering, also called smoothing. This step isperformed to increase the signal-to-noise ratio. Noise is present in fMRI data, but follows a (mostly)Gaussian distribution with an average value of zero.24 As the signal due to neuronal activity is onaverage non-zero, the functional data will be spatially filtered using a Gaussian kernel. The kernelwidth is chosen to be 6 mm, a common choice in literature [30], [31]. The results of the spatialsmoothing are shown in figure 5.5. This step is performed using the smooth function from SPM, whichis used by the CONN toolbox [38], [37].

(a) Functional data before spatial smoothing. (b) Functional data a�er spatial smoothing.

Figure 5.5: Visualization of the e�ect of spatial smoothing.

24. h�p://mindhive.mit.edu/node/112

29

Parcellation

5.2 Parcellation

5.2.1 Human brain atlas

Both the intensity features and the connectivity features use a human brain atlas. A human brainatlas is a representation of the human brain where the brain is subdivided in several brain regions, alsocalled parcels. The Lausanne brain atlas is used [50]. The Lausanne brain atlas is defined in the MNIspace (see section 5.1.7) and has five di�erent levels of brain region division. The first level divides thewhole brain in 89 parcels, the second level has a division of 129 parcels, the third level has a division of234 parcels, the fourth level has a division of 463 levels and the fi�h level has a division of 1015 parcels.

Both the intensity and connectivity features will be calculated for each atlas level separately. Thisis done to investigate the influence of parcel size on the resulting features.

5.2.2 Human brain atlas resizing

The preparatory step that is taken is the resizing of the Lausanne brain atlas. The Lausanne brainatlas has been built using structural MRI images and has a higher resolution than the fMRI data fromthe data set. All atlas levels will be resized to match the fMRI data using the imresize3 function fromMATLAB.25 The nearest neighbor option for interpolation is used.

5.3 Intensity features

The first category of features that will be explored are intensity features. They reflect the averageactivity of the di�erent brain regions during resting state and are obtained from the fMRI data. Itshould be noted that these features only give a simplified representation of the brain activity as notall information in the fMRI data is used.

5.3.1 Absolute intensity and relative intensity

Two intensity feature subtypes will be explored: absolute and relative intensity features. The absoluteintensity features are the features that reflect the average intensity of each brain region. They arefound by averaging all time series to a single value for each time series and then averaging all valuesof a single brain region that is defined by the Lausanne brain atlas. It should be noted that this type offeature is sensitive to di�erences between patient scans. A global elevation or decrease of the recordedvalues within the scan of a patient will influence the final feature values of the patient significantly.Relative intensity features counter this problem.

Relative intensity features also reflect the average activity of brain regions through time. The maindi�erence with absolute intensity features is the fact that the intensity of the di�erent brain regionsis normalized for each patient personally. This is done using formula 5.2. Here Irelative,j is the relativeintensity feature value of brain region j, Iabsolute,j is the absolute intensity feature value of brain regionj and Ibrain,average is the average intensity value of the whole brain. Ibrain, average is calculated byaveraging all values of all time series.

Irelative,j =Iabsolute,j − Ibrain,average

Ibrain,average(5.2)

25. h�ps://www.mathworks.com/help/images/ref/imresize3.html

30

Functional features

5.3.2 Feature selection process

The feature selection process is shown in figure 5.6. The process starts with the preprocessed data (seesection 5.1) and can be divided in two parts: a patient specific and a group specific part. The patientspecific part defines the steps that are done on the data of each patient separately, the group specificpart defines the steps that are done on a group level (healthy controls and depression patients). Eachdi�erent step will be explained.

Figure 5.6: The intensity feature selection process.

5.3.2.1 Time averaging

The first step of the patient specific part of the feature selection process is time averaging. The timeseries of the fMRI data of each patient will be averaged. This reduces the patient data matrix from a30076x160 or 30076x290 matrix (30076 equals the amount of time series that overlap with the Lausannebrain atlas, 160 or 290 equals the amount of di�erent time points in an fMRI time series) to a 30076x1matrix.

5.3.2.2 Parcel averaging

The second step of the patient specific part is parcel averaging. The 3D location of each time series(now average value) is mapped to the Lausanne brain atlas and linked to the corresponding parcel. Thedi�erent time series that belong to the same parcel are averaged, resulting in a single average intensityvalue for each parcel. This reduces the patient data matrix from a 30076x1 matrix to a Nx1 matrix (Nequals the amount of parcels of each atlas level, see section 5.2.1). As five atlas levels exist and bothabsolute and relative intensity values are calculated, ten di�erent data matrices are calculated for eachpatient.

31

Intensity features

5.3.2.3 Group averaging

The third step of the feature selection process and first step of the group specific part is group averag-ing. The value of each parcel is averaged for both groups (healthy controls and depression patients).This results in 10 Nx1 matrices for each group.

5.3.2.4 Di�erence calculation

The fourth step of the feature selection process is di�erence calculation. The average group values ofthe depression patients are subtracted from the average group values of the healthy controls. Thisresults in ten Nx1 matrices called di�erence matrices.

5.3.2.5 Feature selection

The final step is the feature selection itself. The 20 parcels that have the biggest absolute value in thedi�erence matrices are selected. These parcels have the biggest di�erent intensity values on averagebetween the healthy controls and depression patients and are therefore best suited to be used asfeatures to distinguish both groups. The significance of the features was validated statistically usingthe two-sample t-test, done by the �est2 function from MATLAB.26

As five atlas levels are used and two feature subtypes are calculated, ten feature sets are obtained.Each feature set is a 106x21 matrix; 106 equals the total amount of people in the data set and 21equals the twenty features plus a final value defining the group to which the person belongs. Allfeature sets are normalized using the z-score.

26. h�ps://www.mathworks.com/help/stats/�est2.html

32

Functional features

5.3.3 Interpretation of the found features

As five atlas levels are used and two feature subtypes are calculated, ten feature sets are obtained. Twosets will be discussed in this section: the best performing feature set of each intensity feature subtype.The results of the classification process using these two feature sets can be found in section 7.3.1, thediscussion of the results in section ??.

5.3.3.1 Interpretation of the used tables

The Number column defines the significance of the feature. Feature 1 defines the parcel that has thebiggest value in the di�erence matrix, feature 20 defines the parcel that has the twentieth biggestdi�erence value in the di�erence matrix (see section 5.3.2.4). The Parcel column defines the nameof the parcels that are related to the features. These are the parcels whose feature values di�er onaverage the most between the healthy controls and the depression group. "(L)" denotes that the parcelis located in the le� hemisphere, "(R)" denotes that the parcel is located in the right hemisphere. Thenumber at the end of the parcel name is its further subdivision number (brain regions are subdivided insmaller regions when higher atlas levels are used). The Sign column defines the sign of the di�erencevalue. A "+" sign denotes a positive di�erence value, meaning that the intensity value of the parceland by extension the general activity of that parcel is higher in the healthy controls compared to thedepression group. A "-" sign defines the contrary.

5.3.3.2 Absolute intensity features

The twenty most significant features are shown in table 5.1. A first noticeable fact is that severalfeatures are subdivisions of a single brain region. Five features (feature 1, 9, 12, 18 and 20) area subdivision of the le� superior frontal gyrus (LSFG), three features (feature 3, 6 and 11) are asubdivision of the right superior frontal gyrus (RSFG), two features (feature 4 and 15) are a subdivisionof the right rostral middle frontal gyrus (RRMFG) and six features (feature 5, 7, 8, 10, 13 and 14) area subdivision of the le� rostral middle frontal gyrus (LRMFG). It should be noted that all but one ofthese features (feature 7) are denoted by a "+" sign, meaning that all but one of these regions showhypoactivity in the depression group compared to the healthy controls. This shows that hypoactivityin the frontal lobe is clearly present in both hemispheres in this data set. This phenomenon has beenreported in literature [51]. A second observation that is made is that the four brain regions that containmultiple features are two corresponding brain regions lying in both hemispheres, showing again thestrong localization of (significant) hypoactivity to the frontal regions in the depression group whencompared to the healthy controls. The di�erence in feature amount between these regions with respectto the hemisphere they are located in could point to the frontal asymmetry commonly found in peoplewith depression [52], [53]. This di�erence in feature amount between the le� and right hemisphere isalso observed with the structural features (see section 4.2.2.2).

All four brain regions (LSFG, RSFG, LRMFG, RRMFG) are located in the prefrontal cortex. Hypoac-tivity in the prefrontal cortex is common in people with depression and has been described multipletimes [54], [55], [56]. The LSFG specifically is a very significant brain region related to depression asit is one of the main regions for transcranial magnetic stimulation (TMS). The first (and therefore mostsignificant) feature that is defined is the third subdivision of the LSFG and is located at the borderbetween Brodmann area 10 and 46; this location almost exactly matches one of the locations for TMSfor patients with depression [57], [58].

Another observation that can be made is the fact that both frontal poles are features that, contrary toalmost all features, have a negative sign. This means that both frontal poles show increased activityin the depression group when compared to the healthy controls. This seems counter-intuitive as the

33

Intensity features

prefrontal cortex, which contains the frontal poles, is hypoactive in people with depression [54]. Noclear explanation for this is currently known.

It can be concluded that, from a clinical relevance point of view, the absolute intensity features arehighly significant.

Table 5.1: Absolute intensity features (Atlas3).

Number Parcel Sign1 Superior frontal gyrus 3 (L) +2 Frontal pole (L) -3 Superior frontal gyrus 3 (R) +4 Rostral middle frontal gyrus 1 (R) +5 Rostral middle frontal gyrus 4 (L) +6 Superior frontal gyrus 4 (R) +7 Rostral middle frontal gyrus 6 (L) -8 Rostral middle frontal gyrus 2 (L) +9 Superior frontal gyrus 5 (L) +10 Rostral middle frontal gyrus 3 (L) +11 Superior frontal gyrus 2 (R) +12 Superior frontal gyrus 2 (L) +13 Rostral middle frontal gyrus 1 (L) +14 Rostral middle frontal gyrus 5 (L) +15 Rostral middle frontal gyrus 3 (R) +16 Frontal pole (R) -17 Caudal middle frontal gyrus (L) +18 Superior frontal gyrus 4 (L) +19 Precentral gyrus (R) +20 Superior frontal gyrus 9 (L) +

5.3.3.3 Relative intensity features

The twenty most significant features are shown in table 5.2. A similar observation as with the absoluteintensity features (see section 5.3.3.2) is made; most features are subdivisions from a few larger brainregions. The larger brain regions are the same as the brain regions found with the absolute intensityfeatures. Seven features (feature 1, 2, 4, 6, 9, 11 and 15) are a subdivision from the le� superior frontalgyrus (LSFG), four features (feature 7, 10, 12 and 20) are a subdivision from the le� rostral middlefrontal gyrus (LRMFG), three features (feature 3, 8 and 13) are a subdivision of the right superiorfrontal gyrus (RSFG) and two features (feature 14 and 19) are a subdivision of the right rostral middlefrontal gyrus (RRMFG). The clinical significance of these features (as well as the le� frontal pole,feature 5) has been discussed in the previous section (section 5.3.3.2) and will not be repeated.

The absolute intensity and relative intensity features are largely located in the same brain regions. Thisis expected, but strengthens the validity of the absolute intensity features. While the relative intensityfeatures reflect the relative activity of a single brain region with respect to the global average activity,absolute intensity features only reflect absolute activities (see section 5.3.1). The similar features showthat the absolute intensity features do not su�er from the lack of a normalization. It should be notedthat this conclusion is only valid for the used data set. Any extension of this feature type to other datasets should be validated with a similar absolute versus relative intensity comparison. A di�erencebetween the absolute and relative intensity features is the presence of two features (feature 14 and19), located in the RRMFG, that have a negative sign. No clear explanation for this is currently known.

34

Functional features

Table 5.2: Relative intensity features (Atlas5).

Number Parcel Sign1 Superior frontal gyrus 23 (L) +2 Superior frontal gyrus 21 (L) +3 Superior frontal gyrus 8 (R) +4 Superior frontal gyrus 12 (L) +5 Frontal pole (L) -6 Superior frontal gyrus 4 (L) +7 Rostral middle frontal gyrus 10 (L) +8 Superior frontal gyrus 15 (R) +9 Superior frontal gyrus 9 (L) +10 Rostral middle frontal gyrus 1 (L) -11 Superior frontal gyrus 39 (L) +12 Rostral middle frontal gyrus 25 (L) -13 Superior frontal gyrus 37 (R) +14 Rostral middle frontal gyrus 1 (R) -15 Superior frontal gyrus 19 (L) +16 Precentral gyrus (R) +17 Postcentral gyrus 15 (R) +18 Postcentral gyrus 10 (R) +19 Rostral middle frontal gyrus 26 (R) -20 Rostral middle frontal gyrus 5 (L) +

35

Connectivity features

5.4 Connectivity features

The second type of features that will be explored are connectivity features. These features reflectdi�erences in functional connectivity between the healthy controls and the depression group. Func-tional connectivity is a general term describing di�erent techniques that evaluate the functionalconnection between di�erent brain regions that might be spatially separated [4], [59]. The mostcommon connectivity measures are correlation, cross-correlation (CR), coherence, granger causality(GCI), directed transfer function (DTF), partial directed coherence (PDC) and mutual information (MI)[4].

5.4.1 Functional connectivity measures

The functional connectivity measures di�er from each other and thus reflect di�erent aspects of thefunctional connection. The properties most commonly used to distinguish the measures are whetherthey are directed or undirected, whether they are bi- or multivariate, whether they operate in thetime- or frequency domain and whether they are linear or non-linear. The properties of the previouslymentioned connectivity measures are shown in table 5.3

Undirected vs directed

An undirected connectivity measure can only test whether a functional connection exists betweenbrain regions. Directed connectivity measures can, aside from the validation of a possible connection,also determine in which way the connection exists. This direction in the connection gives extrainformation as it appoints a role to the di�erent brain regions: one brain region has influenced theother brain region. The direction of a functional connection can be used to explore how signalspropagate in the brain.

Bivariate vs multivariate

Bivariate connectivity measures can only analyze the connection of two di�erent brain regions at atime. More complex relationships between multiple brain regions (e.g. brain networks such as thedefault mode network) can be documented using multivariate connectivity measures.

Time domain vs frequency domain

Time domain connectivity measures will analyze if a connection exists between brain regions in thetime domain while frequency domain measures analyze possible connections in the frequency domain.

Linear vs non-linear

Linear connectivity measures will analyze if a linear relationship exists between brain regions whilethe non-linear connectivity measures also analyze if non-linear relationships are present.

Table 5.3: Functional connectivity measures [4].

Measure Un-/directed Bi-/multivariate Time/Frequency Non-/LinearCorrelation undirected bivariate time linearCR directed bivariate time linearCoherence undirected bivariate frequency linearGCI directed bivariate time linearDTF directed multivariate frequency linearPDC directed multivariate frequency linearMI undirected bivariate time non-linear

36

Functional features

5.4.1.1 Selection process of connectivity measures

The selection of connectivity measures is, as explained in section 5.4.1, best done using the distinguish-ing properties. Aside from the properties it is important to keep in mind what the final goal of theconnectivity measures is: the classification of people in two possible groups. Using this consideration,a conclusion can be drawn which is that the final result of a connectivity measure should be a singlevalue, or should be easily reducible to a single value, as this is needed for classification.

Frequency based connectivity measures are not investigated. Due to the limited frequency range([ 1

128 , 0.5] Hz) imposed by the high frequency filter (see section 5.1.6) and the limitation of the EPIsequence (see section 2.1.3.1), the frequency spectrum of fMRI is of li�le interest for feature selectionwhen compared to EEG data.

A second reduction is done by eliminating all directed features. This reduction is defended by the factthat fMRI measures the hemodynamic response, which has variable time delays throughout the brain.Due to this variability, directed connectivity measures are di�icult to interpret. A second validationfor this reduction is the fact that previous a�empts at depression classification based on resting statefMRI used undirected methods [30], [31].

Using both reductions, two connectivity measures are selected: correlation and mutual information.Both will be explained in detail below.

Correlation

Correlation, the Pearson correlation coe�icient more precisely, is a measure of the linear correla-tion between two signals. It has a possible range of [-1,1] where -1 means there is a total negativecorrelation between both signals, 1 means there is a total positive correlation between both signalsand 0 means no correlation exists between the signals [60]. Although correlation might seem a simpleconnectivity measure, it is proven that it can perform equally as well as mutual information in correctlydefining linear connections and is commonly used in functional connectivity studies [30], [31], [61].

The calculation of the Pearson correlation coe�icient of two signals, A and B, is shown in formula 5.3.Here N denotes the amount of di�erent values both signals A and B consist of, Ai, µA and Bi, µB arethe ith sample and mean value of variable A and B respectively. The MATLAB function corrcoef wasused to calculate the Pearson correlation coe�icient.27

ρ(A,B) =1

N − 1

N∑i=1

(Ai − µAσA

)(Bi − µBσB

)(5.3)

Mutual information

Mutual information is a non-linear connectivity measure which quantifies the mutual dependencyof two signals. It calculates how much information can be obtained of the second variable whenonly the first variable is observed [62]. The possible range of values is [0,+∞[, where a higher valuedenotes more shared information between both signals. Because this range is unbounded, the onlyinformation one can extract is the di�erence between two values [62]. Formula 5.4 represents thecalculation process. Here pAB(a, b) denotes the joint probability between signal A and B, calculatedusing their combined histogram; pA and pB denote the probability of signal A and B respectively,

27. h�ps://www.mathworks.com/help/matlab/ref/corrcoef.html

37

Connectivity features

these can again be calculated using their histograms. A MATLAB package28 is used to calculate themutual information.

MIAB =∑a,b

pAB(a, b)log

(pAB(a, b)

pA(a)pB(b)

)(5.4)

5.4.2 Feature selection process

The feature selection process is shown in figure 5.7. It is quite similar to the feature selection processof the intensity-based features (see section 5.3). It starts with the preprocessed data, as described insection 5.1, and can be divided in two parts: a patient specific part and a group specific part. Thepatient specific part of the feature selection process contains the preparation of the data for eachpatient while the group specific part contains the steps performed on a group level (depression patientsand healthy controls). Each step will be explained.

Figure 5.7: The connectivity feature selection process.

5.4.2.1 Global signal regression

The principle of global signal regression is the idea that certain processes, such as cardiac pulse andbreathing, influence the BOLD signals (see section 2.2.1) captured by the MRI scanner. These processeswill influence the BOLD signal throughout the whole brain by introducing an extra signal that iscommon in every time series captured. The captured signal is the addition of two signals: a globalsignal and the local signal caused by local neuronal changes. The addition of a common signal in

28. h�ps://www.mathworks.com/matlabcentral/fileexchange/13289-fast-mutual-information-of-two-images-or-signals

38

Functional features

every time series leads to an artificially increased connectivity between every voxel. This artificialincrease leads to falsely inflated connectivity values. This increase in connectivity throughout thebrain is unwanted as only the neuronal changes within the brain are a true indication of connectivitybetween brain regions.

To counter the introduction of a global signal, global signal regression is applied. Global signal re-gression calculates the global signal by averaging all time series and uses the average signal as aregressor in a general linear model. The general linear model regresses the global signal out of thetime series, which are now assumed to only contain the information regarding the neuronal changeswithin the brain. As only the neuronal changes are present in the time series, connectivity values nowreflect true connectivity between brain regions [63], [64].

Global signal regression is however subject of much discussion within the field of neuroscience. Theprincipal reasoning behind global signal regression is correct, but an assumption is made. The as-sumption is the idea that the influences inducing the global signal and the local neuronal changesare completely unrelated (they are orthogonal to each other), and that by regressing the global signalout, one does not regress part of the neuronal signal out. Multiple studies have concluded that globalsignal regression actually introduces anti-correlated networks and therefore might itself introducefalse correlations [65], [66], [67]. This makes the use of global signal regression a controversial choice.

As no definite conclusion is reached regarding global signal regression, connectivity features will becalculated with and without global signal regression as a preprocessing step (as indicated by the do�edline in figure 5.7). The data set created without the global signal regression step will be called the non-regressed data set, the data set created with the global signal regression step will be call the regresseddata set. The CONN toolbox is used to perform global signal regression [38].

5.4.2.2 Parcel simplification

A second step that is performed is parcel simplification. This step will reduce all time series froma parcel into a single time series. This step is necessary for two reasons: dimension reduction andintrapatient/intergroup voxel variations.

Dimension reduction is needed as every patient has 155648 (64x64x38) unique time series, 30076time series which are overlapping with the Lausanne brain atlas [50]. If no dimension reduction isapplied, a connectivity measure matrix with size 30076x30076 is obtained, which is computationallynot achievable.

The intrapatient/intergroup variations, meaning voxel variability both within a single patient scanand between di�erent patient scans, make single voxel time series highly unreliable as a basis forfeatures.

The technique used to perform the parcel simplification is principal component analysis (PCA). PCAis a statistical technique in which a group of possible correlated observations (here time series) aretranslated into a set of new variables that are not correlated anymore (they are orthogonal to eachother). Each variable in the new set is called a principal component [68]. Only the first principalcomponent of each parcel will be used to calculate functional connectivity between parcels as it willcompensate for as much variance as possible.

The result from the parcel simplification step is a 1xN matrix for each patient, where N is 83, 129,234, 463 or 1015 depending on the chosen atlas level (see section 5.2.1). Each cell contains the first

39

Connectivity features

principal component of the corresponding parcel. Principal component analysis is performed using aMATLAB toolbox.29

5.4.2.3 Connectivity measure calculation

The next step in the feature selection process is the connectivity measure calculation. The two connec-tivity measures that are used are correlation and mutual information (see section 5.4.1.1). The resultof the connectivity measure calculation is an NxN matrix for each patient, where N is 83, 129, 234,463 or 1015 depending on the chosen atlas level (see section 5.2.1). Cell Ci,j contains the calculatedconnectivity between parcels i and j. These matrices are symmetric because undirected connectivitymeasures are used, meaning that Ci,j is equal to Cj,i. It should be noted that mutual information forN = 463, 1015 is not calculated as the computation time is too long.

5.4.2.4 Group averaging

The next step is group averaging. An average matrix is calculated for both groups (depression patientsand healthy controls). The result of this step is a group matrix for each group, for each atlas level andfor each connectivity measure.

5.4.2.5 Di�erence calculation

A�er the calculation of the group matrices, the depression group matrices are subtracted from thehealthy controls group matrices in order to obtain the di�erence matrices.

5.4.2.6 Feature selection

The last step is the selection of the features. This is done by finding the 20 biggest values for eachdi�erence matrix. These values reflect the connectivity measures that on average di�er the mostbetween both groups and thus will be used to di�erentiate between both groups. The significanceof the features is validated statistically using the two-sample t test, done by the �est2 function fromMATLAB. The features found for mutual information calculated from the non-regressed data set arenot significant. All possible features from this feature subtype are tested using the t-test, none arestatistically significant. Classifiers trained on this feature set do not perform well (see section 7.3.1).

As five atlas levels for correlation and three atlas levels for mutual information are used and two datasets are used (regressed and non-regressed), sixteen feature sets are obtained.

29. h�ps://www.mathworks.com/matlabcentral/fileexchange/38300-pca-and-ica-package

40

Functional features

5.4.3 Interpretation of the found features

A total of 16 feature sets, each containing 20 features, have been calculated. Four sets will be discussedin this section: the best performing feature sets of each connectivity measure for both the regressedand non-regressed data set. The results are shown in section 7.3.1 and discussed in secton ??.

Interpretation of the used tables

The number column defines the significance of the feature. Feature 1 defines the connection betweentwo parcels that has the biggest value in the di�erence matrix (see section 5.4.2.5), meaning that thisconnection has on average the highest di�erence between the healthy controls and the depressiongroup, feature 20 the twentieth biggest di�erence. The region1 and region2 columns define the twoparcels that are involved in the connection the feature reflects. "(L)" denotes that the parcel is locatedin the le� hemisphere, "(R)" denotes the right hemisphere. The number at the end of the region nameis its further subdivision number (brain regions are subdivided in smaller regions when higher atlaslevels are used). The sign column defines the sign of the di�erence value. A "+" denotes a positivedi�erence value, meaning that the connection between the two parcels is higher with the healthycontrols when compared to the depression group. A "-" sign denotes the contrary.

5.4.3.1 Features calculated with the non-regressed data set

Correlation

The twenty most significant features are shown in table 5.4. Feature 13, 16 and 20 are closely linked todepression. The connections reflected by feature 13 and 16 both contain the anterior cingulate cortex(ACC), which is an important factor in several models of depression [69], [70], feature 20 contains theorbitofrontal cortex, which has been discussed in section 5.3.3.2. As this feature set did not performas well as the correlation with regressed data feature set, it will not be discussed further. The resultscan be found in section 7.3.1, the discussion of the results in section ??.

Table 5.4: Correlation with non-regressed data set features (Atlas2).

Number Region1 Region2 Sign1 Inferior temporal gyrus (L) Middle temporal gyrus (L) +2 Superior frontal gyrus (R) Precentral gyrus (L) +3 Precentral gyrus (L) Postcentral gyrus (L) +4 Caudal middle fontal gyrus (R) Rostral middle frontal gyrus (R) +5 Precentral gyrus (R) Inferior temporal gyrus (R) +6 Precentral gyrus (R) Paracentral gyrus (L) +7 Middle temporal gyrus (R) Banks of the superior temporal sulcus (L) -8 Pars triangularis (L) Pars opecuaris (L) +9 Pericalcarine (R) Lingual gyrus (L) -10 Postcentral gyrus (R) Superior temporal gyrus (R) +11 Lateral orbitofrontal cortex (L) Insula (Le�) +12 Precentral gyrus (R) Superior temporal gyrus (R) +13 Lateral orbitofrontal cortex (L) Caudal anterior cingulate cortex (L) +14 Superior temporal gyrus 1 (R) Superior temporal gyrus 3 (R) -15 Caudal middle fontal gyrus (R) Precentral gyrus (R) +16 Rostral anterior cingulate cortex (R) Lateral orbitofrontal cortex (L) +17 Superior frontal gyrus (R) Pars opercularis (L) +18 Precentral gyrus 3 (L) Precentral gyrus 1 (L) +19 Precentral gyrus (R) Supramarginal gyrus (R) +20 Lateral orbitofrontal cortex 3 (R) Lateral orbitofrontal cortex 1 (R) -

41

Connectivity features

Mutual information

The twenty most significant features are shown in table 5.5. All features reflect a connection in whichthe right or le� superior frontal gyrus is involved. Seven features (feature 2, 7, 8, 11, 13, 16 an 18)reflect a connection within the le� or right superior frontal cortex itself. The accuracy results of thisfeature set are lower than 50% (see section 7.3.1.7), so these features will not be discussed further.

Table 5.5: Mutual information with non-regressed data set features (Atlas3).

Number Region 1 Region 2 Sign1 Superior frontal gyrus 6 (R) Superior frontal gyrus 8 (L) -2 Superior frontal gyrus 6 (R) Superior frontal gyrus 7 (R) -3 Superior frontal gyrus 8 (R) Superior frontal gyrus 8 (L) -4 Superior frontal gyrus 8 (L) Superior frontal gyrus 7 (R) -5 Superior temporal gyrus 4 (R) Superior frontal gyrus 8 (L) -6 Superior frontal gyrus 6 (R) Lateral orbitofrontal cortex (R) -7 Superior frontal gyrus 6 (R) Superior frontal gyrus 8 (R) -8 Superior frontal gyrus 8 (L) Superior frontal gyrus 6 (L) -9 Caudal middle frontal gyrus 3 (L) Superior frontal gyrus 8 (L) -10 Superior frontal gyrus 6 (R) Caudal middle frontal gyrus 3 (L) -11 Superior frontal gyrus 7 (L) Superior frontal gyrus 8 (L) -12 Superior frontal gyrus 6 (R) Superior frontal gyrus 6 (L) -13 Superior frontal gyrus 6 (R) Superior frontal gyrus 4 (R) -14 Superior frontal gyrus 8 (L) Lateral occipital sulcus 1 (L) -15 Superior frontal gyrus 8 (L) Lateral orbitofrontal cortex (R) -16 Superior frontal gyrus 4 (R) Superior frontal gyrus 8 (L) -17 Superior frontal gyrus 6 (R) Superior temporal gyrus 4 (R) -18 Superior frontal gyrus 8 (R) Superior frontal gyrus 7 (R) -19 Superior frontal gyrus 6 (R) Lateral occipital cortex 1 (L) -20 Rostral middle frontal gyrus 5 (L) Superior frontal gyrus 8 (L) -

5.4.3.2 Features calculated with the regressed data set

Correlation

The twenty most significant features are shown in table 5.6. Several features are linked to depression,each significant feature will be discussed.

Feature 2 reflects the connection between the right medial orbitofrontal cortex and the le� anteriorcingulate cortex. Connections between the frontal cortex and the anterior cingulate cortex are dis-turbed in depression [55], [71]. Feature 5 reflects the connection between the le� and right precuneus.The precuneus is a part of the default mode network (DMN) and disturbances in the default modenetwork have been reported in people with depression [72], disturbances in this specific connectionbetween the le� and right precuneus in people with depression are however not found in literature.Feature 9 reflects the connection between the precuneus and the isthmus cingulate cortex, whichcould also be related to depression. No literature for this specific connection has been found. Feature13 and 15 are two connections in the prefrontal cortex, which is involved in depression [54], [56]. Thesign of both connections is negative, which is counter-intuitive as the prefrontal cortex is hypoactivein depression [55]. A possible explanation for this could be that hypoactivity does not mean a decreasein correlation between brain regions. No literature for these specific connections has been found.

42

Functional features

Table 5.6: Correlation with regressed data set features (Atlas3).

Number Region 1 Region 2 Sign1 Rostral middle frontal gyrus (L) Supramarginal gyrus (L) +2 Medial orbitofrontal cortex (R) Rostral anterior cingulate cortex (L) +3 Fusiform gyrus (R) Inferior temporal gyrus (R) -4 Lingual gyrus (R) Pericalcarine gyrus (L) +5 Precuneus (R) Precuneus (L) +6 Precentral gyrus (R) Postcentral gyrus (L) +7 Paracentral gyrus 1 (L) Paracentral gyrus 2 (L) -8 Precentral gyrus (R) Precentral gyrus (L) -9 Isthmus cingulate cortex (L) Precuneus (L) +10 Fusiform gyrus Lateral occipital cortex (L) -11 Lateral occipital sulcus (L) Fusiform gyrus (L) +12 Pericalcarine gyrus (R) Pericalcarine gyrus (L) +13 Superior frontal gyrus 6 (R) Superior frontal gyrus 7 (R) -14 Superior parietal gyrus (R) Lateral occipital cortex (R) +15 Superior frontal gyrus (R) Caudal middle frontal gyrus (R) -16 Precentral gyrus (R) Supramarginal gyrus (R) +17 Lateral occipital cortex (R) Fusiform gyrus (R) +18 Superior temporal gyrus (R) Precentral gyrus (L) -19 Superior parietal gyrus 3 (L) Superior parietal gyrus 4 (L) -20 Cuneus (R) Cuneus (L) -

Mutual information

The twenty most significant features are shown in 5.7. A first observation is that a few brain regionsreturn in several features. The four brain regions present in multiple features are the right paracentrallobule, the right inferior parietal gyrus, the brainstem and the right superior parietal gyrus. The rightparacentral lobule (RPL) is present in 11 features (feature 1, 2, 3, 7, 8, 9, 10, 13, 16, 18 and 19), the rightinferior parietal gyrus (RIPG) is present in 9 features (feature 2, 4, 5, 6, 7, 12, 15, 17 and 20), the rightsuperior parietal gyrus (RSPG) is present in four features (feature 1, 5, 11 and 14) and the brainstemis present in three features (feature 3, 4 and 11).

Changes in connectivity in the paracentral lobule in people with depression have been described inliterature [73], [74], [75]. No clear explanation of its involvement in depression is given and di�er-ent literature also reports both increased and decreased functional connectivity within the region.Further research needs to be done to investigate this region. The right inferior and superior parietallobe have also been linked with depression, however decreased functional connectivity is reportedwhile increased connectivity is found here [76]. The brainstem contains multiple nuclei that could beinvolved in depression and research tries to understand its involvement in the disease. The specificconnections involving the brainstem however are not found in literature.

The assessment of the clinical relevance of connectivity features is more di�icult than the assessmentof the intensity or structural features as two regions are involved, making a possible recurrence of thespecific connection in literature less likely. An a�empt to counter this problem is the dismissal of thesubdivision number when possible (subdivision numbers are still shown when a connection within asingle region is found), but this did not improve the search for literature significantly. As di�erentconnectivity measures are used, a recurrence of a connection in literature is not always proof of itsclinical significance. Some skepticism toward the found literature is also necessary; this skepticismis needed for every found relevance for every feature type and is not bounded to the connectivityfeatures.

43

Connectivity features

Table 5.7: Mutual information with regressed data set features (Atlas3).

Number Region 1 Region 2 Sign1 Paracentral lobule (R) Superior parietal gyrus (R) -2 Inferior parietal gyrus (R) Paracentral lobule (R) -3 Paracentral lobule (R) Brainstem -4 Brainstem Inferior parietal gyrus (R) -5 Inferior parietal gyrus 5 (R) Superior parietal gyrus (R) -6 Inferior parietal gyrus 1 (R) Inferior parietal gyrus 5 (R) -7 Inferior parietal gyrus 1 (R) Paracentral lobule (R) -8 Paracentral lobule (R) Parahippocampal gyrus (L) -9 Paracentral lobule (R) Supramarginal gyrus (R) -10 Rostral middle frontal gyrus (R) Paracentral lobule (R) -11 Superior parietal gyrus (R) Brainstem -12 Inferior parietal gyrus (R) Parahippocampal gyrus (L) -13 Fusiform gyrus (R) Paracentral lobule (R) -14 Parahippocampal gyrus (L) Superior parietal gyrus (R) -15 Inferior parietal gyrus (R) Lateral occipital cortex (L) -16 Paracentral lobule (R) Superior temporal gyrus (L) -17 Supramarginal gyrus (R) Inferior parietal gyrus (R) -18 Rostral middle frontal gyrus (L) Paracentral lobule (R) -19 Lateral occipital cortex (L) Paracentral lobule (R) -20 Inferior parietal gyrus (R) Fusiform gyrus (R) -

44

Chapter 6

Classifier training

A pipeline is designed to train each classifier, this to be able to compare results between di�erentfeature sets (see chapter ??). The complete classification training pipeline is shown in figure 6.1. Everystep of this pipeline will be explained.

6.1 Classification training pipeline

6.1.1 Starting point

The starting point of the classifier training pipeline is a feature set, the final result from one of thefeature selection processes (see chapter ??). The feature set is a 106x(N+1) matrix, where N equals theamount of features. The number 106 denotes the amount of people that are present in the whole dataset (60 healthy controls and 46 patients with depression). The last column of the feature set containsthe class of each patient. One denotes the healthy control class, minus one the depression class.

6.1.2 Class balancing

The size of both classes is unequal. This can pose a problem in the training of a classifier becausea bias could be introduced towards the more abundant class. If class A is ten times more abundantthan class B, a machine learning technique that always predicts class A, even without being trained,will have an accuracy of over 90%; this is called the accuracy paradox. Two main possibilities exist tocounter this problem: class imbalance learning methods and class balancing. Class imbalance learningmethods refer to adaptions of learning methods that can counter the problem of class imbalance [88],[89].

Class balancing refers to the dismissal of instances from one class so that both classes are representedequally. This method has a disadvantage when compared to class imbalance learning methods: thedismissal of instances makes the training set smaller, which might lead to overfi�ing30. Furthermoreis class balancing not possible if the imbalance between both classes is too big.

Class balancing is chosen as a balancing method because the dismissed persons (fourteen healthycontrols) can be used as a second validation set (see section 6.1.4.2). The second validation set canbe used to check if the trained model is not overfit. Each time the classification training pipeline isused, the fourteen healthy controls that will not be used for the training of the classifier are chosen byrandom permutation. This ensures that all healthy controls can be used in the training of the classifier.

30. h�ps://www.investopedia.com/terms/o/overfi�ing.asp. Overfi�ing: a modeling error which occurs when a function is tooclosely fit to a limited set of data points.

45

Classification training pipeline

6.1.3 Train-validation spli�ing

The 46 healthy controls that are selected are, together with the 46 patients with depression, the dataset (92 people in total) that is used to train and validate the classifier. The first step is to split this dataset in a training set and a validation set. A common choice is the 80-20 split [44]. 80 percent of thetotal data set (73 people) are selected by a random permutation to be the training set, 20 percent (19people) are selected to be the validation set.

6.1.4 Model training and validation

The final step in the classification training pipeline is model training and model validation.

6.1.4.1 Training

The support vector machine (see chapter 3) is trained using the training set.

6.1.4.2 Validation

The final step in the classification training pipeline is validation. The trained model will be validatedusing the validation set defined in section 6.1.3. The features of the people in the validation set willbe given to the trained classifier as input. A�erwards the classifier will predict whether the giveninput features belong to a person from the healthy control group or from the patients with depressiongroup. The validation is performed to know if the trained classifier is adequate. A second validationset is used, it contains the fourteen people dismissed in section 6.1.2.

Figure 6.1: The complete classification training pipeline.

46

Chapter 7

Results

In this chapter the found results of the di�erent trained classifiers will be discussed. The first part ofthe chapter will present the results of the best performing classifiers (this is called the global results).The second part presents the results of each feature subtype classifier seperately. The third partpresents the results of the combined feature classifiers Firstly the used method to retrieve data aswell as the visual representation of the results are explained.

7.1 Data collection and presentation

7.1.1 Data collection

Three di�erent feature selection methods are used: the first method lets the algorithm itself choosehow many features are optimal for classification, the second method again lets the algorithm itselfchoose which features are best for classification, but imposes the demand that 10 features need tobe chosen, the third method uses all 20 features of a feature set. Twenty iterations of each featureselection method are used in the machine learning pipeline (see section 6.1) to obtain an averageaccuracy of each method. This reduces the chances of falsely inflated or deflated results for thefeature set. The final trained classifier is validated twice: firstly with the validation set defined bythe train-validation spli�ing step (see section 6.1.3) and secondly with the non-used control groupdefined in the class balancing step (see section 6.1.2). The second validation set will be called theoptional validation set from now on.

The classification pipeline (see section 6.1) will be used 60 times for each feature set. Exceptionsto this are the right hemisphere feature set as it only consists of 11 features (this feature set will beused 40 times: once with the variable amount of features, once with all 11 features) and the parcelvolume feature set as it only consists of 6 features (this feature set will be used 20 times with all 6features). The combined feature sets, shown in section 7.3.2, are used 20 times with all features. Boththe intensity and connectivity feature sets contain 40 features, the structural feature set contains 36features (19 le� hemisphere features, 11 right hemisphere features and 6 parcel volume features).

7.1.2 Presentation

Each classification result will be presented by a table showing the accuracy of both the validation set(19 people, "Val." in the table) as well as the optional validation set (14 people, "Opt." in the table),the numbers are percentages. The mean accuracy and standard deviation of each result distributionare shown at the bo�om of the table. The sensitivity, specificity, positive predictive value (PPV) andnegative predictive value (NPV) of the validation set are also shown. These are not calculated for the

47

Data collection and presentation

optional data set because it is impossible to have a true positive and false negative value. A graphpresents the obtained results as a violin plot (see section 7.1.2.1).

7.1.2.1 Violin plot

A violin plot is an extension of a box plot. It shows beside the mean, median and outliers also theprobability density function of the distribution it represents. The thickness of a violin plot at point xrepresents the probability of x given the represented distribution.31 The violin plots are created using aMATLAB toolbox.32 It should be noted that a simplified version of the violin plot is used; it only showsthe mean and the probability density function. Each figure will contain at least two violin plots: oneblue and one yellow. Blue violin plots show the results of the validation set (19 people), yellow violinplots show the results of the optional validation set (14 people) (see chapter 6).

Violin plots are used as they clearly show di�erences between both classification distributions aswell as the compactness/variability of each distribution. The compactness of a distribution is alsorepresented by the standard deviation in the table. Di�erences in classification distributions betweenthe validation and optional validation set could be a sign of high variability within the feature set or ofoverfi�ing. A slightly lower compactness of the optional validation set is however expected comparedto the validation set, as it contains less people. A low compactness of a result distribution is a sign ofa feature set that is not optimal for classification, as the choice of test set and start conditions of thetraining is the defining factor contrary to the quality of the features itself. High accuracy classifiersin a low compactness result distribution should only be considered as viable for classification if theaccuracy of the iteration is high in both the validation and the optional validation set.

An example of possible feature results is shown in figure 7.1. The two violin plots on the le� showthe result distributions of a high quality feature set. Both have high mean accuracies (denoted bythe black line) and have high compactness as they are small violin plots. No significant di�erence inmean accuracy between both result distributions again shows the consistent ability of the classifiersto correctly diagnose depression. The two center violin plots show a bad result as the le� violin plothas a high mean accuracy and a high compactness while the right violin plot has a low mean accuracyand a low compactness. The big di�erence between both violin plots show that the classifiers are notable to consistently predict depression. The two violin plots on the right show again a bad result asboth result distributions have low mean accuracies and low compactness.

Figure 7.1: Example of the di�erent possible results and violin plots.

31. h�ps://datavizcatalogue.com/methods/violin_plot.html32. h�ps://www.mathworks.com/matlabcentral/fileexchange/45134-violin-plot

48

Results

7.2 Global results

The global results represent the best performing classifier of each feature subtype. The results areshown in figure 7.2. Six feature subtype result distributions are shown: the le� hemisphere thick-ness, parcel volume, absolute intensity (atlas3), relative intensity (atlas3), mutual information withregressed data (atlas3) and correlation with regressed data (atlas3) result distribution. The resultdistributions that are shown are from the validation set and not the optional validation set (hence thecolor of every result distribution). These specific result distributions are chosen as they are the bestsingle feature subtype result distributions (with respect to the mean accuracy and standard deviation).

Two conclusions from the global results can be drawn. The first is that the mean accuracy of thestructural features (le� hemisphere thickness and parcel volume) is much lower than the mean accu-racy of both the intensity and connectivity features. This could possibly be explained by the fact thatthe assumption is made that all people in the depression group have had a similar form of depressionwhile in reality this is not true (this is discussed more in depth in section 8.2). A second conclusionis that while the mean accuracies of the intensity (absolute and relative intensity) and connectivity(mutual information with regressed data and correlation with regressed data) result distributions areclose together, the connectivity result distributions are slightly higher. The higher compactness ofthese result distributions also indicate a be�er feature quality. This could be explained by the factthat the connectivity features extract more information from the fMRI data (this is discussed more indepth in section 8.2).

Figure 7.2: Violin plot of the global results.

49

Results

7.3 Results

The results are discussed for each feature subtype (absolute intensity, relative intensity, correlationwith non-regressed data, mutual information with non-regressed data, correlation with regresseddata, mutual information with regressed data, le� hemisphere thickness, right hemisphere thickness,parcel volume) separately. The results are investigated with respect to the atlas level (for intensity andconnectivity feature sets), the amount of features used (for all feature types) and the compactness ofthe result distributions (for all feature types). Positive outliers are discussed.

7.3.1 Single feature type classifier

7.3.1.1 The le� hemisphere thickness classifier

Description

The results of the le� hemisphere thickness classifier are shown in table 7.1 and figure 7.3. The meanaccuracy remains constant when more features are used for classification, but never reaches levels thatcould be considered adequate for classification as the best mean accuracy is ±59.5%. Compactness islow for all result distributions. One positive outlier is present: fourth iteration when 10 features areused (average accuracy = ±72%, sensitivity = 0.78, specificity = 0.7, ppv = 0.7, npv = 0.78). It should benoted that this is a positive outlier when only the current result distribution is considered.

Figure 7.3: Violin plot of the results of the le� hemisphere thickness classifier.

Table 7.1: Best results of the le� hemisphere thickness classifier.

Iteration Var. features 10 features 19 featuresVal.(%) Opt.(%) Val.(%) Opt.(%) Val.(%) Opt.(%)

1 63,16 42,86 42,11 100 63,16 64,292 78,95 57,14 52,63 64,29 26,32 03 68,42 64,29 63,16 71,43 73,68 42,86

50

Results

4 36,84 28,57 73,68 71,43 21,05 1005 63,16 71,43 63,16 28,57 26,32 06 78,95 57,14 78,95 50 63,16 64,297 31,58 57,14 52,63 57,14 78,95 64,298 57,89 57,14 52,63 35,71 42,11 71,439 57,89 50 47,37 64,29 52,63 85,7110 78,95 57,14 63,16 64,29 68,42 35,7111 57,89 42,86 36,84 57,14 52,63 71,4312 47,37 64,29 57,89 85,71 57,89 71,4313 42,11 85,71 36,84 85,71 68,42 64,2914 57,89 64,29 57,89 42,86 57,89 64,2915 47,37 50 57,89 50 57,89 57,1416 47,37 92,86 63,16 64,29 68,42 64,2917 63,16 64,29 36,84 100 68,42 57,1418 73,68 50 52,63 57,14 68,42 71,4319 36,84 42,86 57,89 57,14 73,68 57,1420 57,89 71,43 57,89 71,43 63,16 57,14Mean accuracy 57,37 58,57 55,26 63,93 57,63 58,21Standard deviation 14,38 14,95 11,39 18,89 16,6 24Sensitivity 0.598 - 0.58 - 0.602 -Specificity 0.549 - 0.525 - 0.551 -PPV 0.595 - 0.55 - 0.575 -NPV 0.605 - 0.556 - 0.578 -

7.3.1.2 The right hemisphere thickness classifier

Description

The results of the right hemisphere thickness classifier are shown in table 7.2 and figure 7.4. Only11 features were statistically significant (see section 4.2.2.2), so the only two di�erent amounts offeatures were used for classification. The mean accuracy does not increase significantly with respectto the amount of features that are used, the best mean accuracy is reached when a variable amount offeatures is used (mean accuracy = ±56.5%). Compactness di�ers significantly between the validationand optional validation set. The reason for this is explained in section ??. One positive outlier ispresent: the tenth iteration when a variable amount of features is used (average accuracy = ±82.3%,sensitivity = 0.8, specificity = 0.78, ppv = 0.8, npv = 0.78). It is di�icult to know whether this outlier isthe product of a good training process or a good starting position.

Table 7.2: Best results of the right hemisphere thickness classifier.

Iteration Var. features 11 featuresVal.(%) Opt.(%) Val.(%) Opt.(%)

1 47,37 71,43 36,84 1002 52,63 57,14 57,89 64,293 52,63 35,71 42,11 04 52,63 28,57 57,89 42,865 52,63 42,86 57,89 506 57,89 35,71 57,89 35,717 63,16 78,57 73,68 57,148 73,68 42,86 57,89 57,149 47,37 64,29 63,16 64,2910 78,95 85,71 47,37 100

51

Results

11 73,68 42,86 63,16 64,2912 68,42 71,43 63,16 35,7113 57,89 28,57 47,37 5014 47,37 57,14 52,63 57,1415 57,89 42,86 57,89 5016 68,42 42,86 57,89 85,7117 63,16 35,71 36,84 10018 42,11 57,14 47,37 21,4319 78,95 64,29 36,84 57,1420 63,16 64,29 36,84 78,57Mean accuracy 60 52,5 52,63 58,57Standard deviation 10,99 16,91 10,66 25,87Sensitivity 0.626 - 0.554 -Specificity 0.574 - 0.498 -PPV 0.595 - 0.535 -NPV 0.605 - 0.517 -

Figure 7.4: Violin plot of the results of the right hemisphere thickness classifier.

7.3.1.3 The parcel volume classifier

Description

The results of the parcel volume classifier are shown in table ?? and figure 7.5. Only 6 features werestatistically significant (see section 4.3) so the classifiers will be trained only with all 6 features. Themean accuracy as well as the compactness di�er significantly between the validation and optionalvalidation set (mean accuracy = ±61%). One positive outlier is present: the twel�h iteration (averageaccuracy = ±71.6%, sensitivity = 0.8, specificity = 0.78, ppv = 0.8, npv = 0.78). It should be noted thatthis is a positive outlier when only the current distribution is considered.

52

Results

Table 7.3: Best results of the parcel volume classifier

Iteration 6 featuresVal.(%) Opt.(%)

1 68,42105 | 35,714292 52,63158 503 52,63158 504 36,84211 42,857145 42,10526 1006 57,89474 507 57,89474 92,857148 47,36842 42,857149 42,10526 10010 36,84211 10011 36,84211 10012 78,94737 64,2857113 52,63158 42,8571414 52,63158 64,2857115 73,68421 57,1428616 26,31579 10017 47,36842 85,7142918 73,68421 57,1428619 47,36842 10020 78,94737 42,85714Mean accuracy 53,15789 68,92857Standard deviation 15,07139 24,98926Sensitivity 0.558 -Specificity 0.504 -PPV 0.535 -NPV 0.528 -

Figure 7.5: Violin plot of the results of the parcel volume classifier.

53

Results

7.3.1.4 The absolute intensity classifier

Description

The results of the absolute intensity classifier are shown in table 7.4 and figure 7.6. The best averageclassification accuracy is obtained when the features were calculated using the third atlas level (234parcels). An increase in classification accuracy is obtained when all 20 features are used for classifi-cation compared to the variable amount of features or 10 features. The compactness of the validationset result distributions is similar (SD = ±10.2) for all three feature amounts and is smaller than thecompactness of the optional validation sets. A significant increase in compactness and mean accuracyis noticeable for the third optional validation set. Two positive outliers are present: the first iterationwhen a variable amount of features is used (average accuracy = ±87%, sensitivity = 0.89, specificity =0.9, PPV = 0.89, NPV = 0.9) and the eleventh iteration when all 20 features are used (average accuracy= ±84%, sensitivity = 0.89, specificity = 0.8, PPV = 0.8, NPV = 0.89).

Table 7.4: Best results of the absolute intensity feature classifier (Atlas3).

Iteration Var. features 10 features 20 featuresVal.(%) Opt.(%) Val.(%) Opt.(%) Val.(%) Opt.(%)

1 89,47 85,71 73,68 78,57 78,95 78,572 63,16 71,43 57,89 64,29 78,95 78,573 57,89 57,14 68,42 50 84,21 57,144 68,42 42,86 68,42 50 68,42 71,435 68,42 78,57 68,42 42,86 42,11 1006 47,37 71,43 52,63 85,71 73,68 71,437 63,16 78,57 68,42 78,57 84,21 78,578 63,16 71,43 78,95 64,29 63,16 78,579 78,95 35,71 94,74 57,14 84,21 64,2910 57,89 50 73,68 42,86 73,68 78,5711 73,68 71,43 57,89 64,29 89,47 78,5712 68,42 71,43 73,68 78,57 84,21 64,2913 63,16 50 73,68 78,57 63,16 85,7114 57,89 64,29 73,68 42,86 68,42 64,2915 73,68 57,14 47,37 85,71 73,68 78,5716 47,37 71,43 68,42 85,71 68,42 92,8617 63,16 42,86 78,95 71,43 73,68 78,5718 63,16 57,14 68,42 85,71 78,95 78,5719 47,37 64,29 57,89 71,43 68,42 57,1420 63,16 14,29 68,42 50 68,42 85,71Mean accuracy 63,95 60,36 68,68 66,43 73,42 76,07Standard deviation 10,16 16,82 10,19 15,34 10,32 10,66Sensitivity 0.67 - 0.712 - 0.762 -Specificity 0.613 - 0.661 - 0.718 -PPV 0.635 - 0.685 - 0.72 -NPV 0.644 - 0.689 - 0.75 -

54

Results

Figure 7.6: Violin plot of the results of the absolute intensity classifier.

7.3.1.5 The relative intensity classifier

Desription

The results of the relative intensity classifier are shown in table 7.5 and figure 7.7. The best averageclassification accuracy is obtained when the features were calculated using the fi�h atlas level (1015parcels). An increase in classification accuracy is obtained when all 20 features are used for clas-sification compared to the variable amount of features or 10 features (mean accuracy = ±73%). Thecompactness of the validation set result distributions varies slightly and is lowest when 10 features areused. No significant increase in compactness is noticed when all 20 features are used. Three positiveoutliers are present: the sixth iteration when a variable amount of features is used (mean accuracy =±81.5%, sensitivity = 0.778, specificity = 0.8, PPV = 0.8, NPV = 0.778), the thirteenth iteration whenall features are used (average accuracy =±84%, sensitivity = 0.9, specificity = 0.889, PPV = 0.889, NPV= 0.9) and the sixteenth iteration when 10 features are used (average accuracy = ±88%, sensitivity =0.889, specificity = 0.8, PPV = 0.889, NPV = 0.8).

Table 7.5: Best results of the relative intensity feature classifier (Atlas5).

Iteration Var. features 10 features 20 featuresVal.(%) Opt.(%) Val.(%) Opt.(%) Val.(%) Opt.(%)

1 52,63 57,14 63,16 78,57 68,42 64,292 57,89 64,29 68,42 100 84,21 71,433 57,89 71,43 52,63 50 73,68 42,864 57,89 57,14 68,42 42,86 68,42 92,865 73,68 64,29 52,63 92,86 73,68 85,716 78,95 85,71 73,68 71,43 89,47 71,437 63,16 42,86 68,42 64,29 73,68 71,438 63,16 71,43 57,89 50 68,42 85,719 63,16 35,71 36,84 85,71 78,95 5010 63,16 64,29 73,68 71,43 57,89 71,43

55

Results

11 63,16 57,14 73,68 85,71 52,63 64,2912 63,16 50 57,89 85,71 89,47 57,1413 57,89 57,14 78,95 64,29 89,47 78,5714 47,37 50 52,63 71,43 68,42 78,5715 52,63 57,14 57,89 28,57 73,68 57,1416 57,89 35,71 84,21 92,86 73,68 85,7117 73,68 85,71 68,42 64,29 73,68 85,7118 63,16 92,86 78,95 71,43 68,42 78,5719 73,68 35,71 57,89 85,71 68,42 78,5720 52,63 35,71 63,16 64,29 78,95 85,71Mean accuracy 61,84 58,57 64,47 71,07 73,68 72,86Standard deviation 7,96 16,69 11,15 17,85 9,415 13,09Sensitivity 0.648 - 0.671 - 0.761 -Specificity 0.59 - 0.619 - 0.714 -PPV 0.605 - 0.64 - 0.73 -NPV 0.633 - 0.65 - 0.744 -

Figure 7.7: Violin plot of the results of the relative intensity classifier.

7.3.1.6 The correlation with non-regressed data classifier

Description

The results of the correlation with non-regressed data classifier are shown in table 7.6 and figure 7.8.The best average classification accuracy is obtained when the features were calculated using thesecond atlas level (129 parcels). The mean accuracy increases consistently when more features areused, the highest mean accuracy is obtained when all 20 features are used (mean accuracy = ±77%).The compactness of the validation set result distributions are similar (SD = ±10.8), the compactnessof the optional validation set fluctuates. The considerable decrease in compactness of the optionalvalidation set result distribution when all 20 features are used (SD = 20.53) is due to both very low

56

Results

and very high accuracies (fi�h and nineteenth iteration, seventh and seventeenth iteration), showingthe unreliability of the complete feature set. Two positive outliers are present: the seventh iterationwhen all 20 features are used (average accuracy = ±86%, sensitivity = 0.8, specificity = 0.778, PPV =0.8? NPV = 0.778) and the seventeenth iteration when all 20 features are used (average accuracy =±90.7%, sensitivity = 0.889, specificity = 0.9, PPV = 0.889, NPV = 0.9).

Figure 7.8: Violin plot of the results of the correlation with non-regressed data classifier.

Table 7.6: Best results of the correlation with non-regressed data classifier (Atlas2).

Iteration Var. features 10 features 20 featuresVal.(%) Opt.(%) Val.(%) Opt.(%) Val.(%) Opt.(%)

1 52,63 64,29 57,89 71,43 68,42 85,712 68,42 42,86 63,16 50 84,21 92,863 31,58 64,29 63,16 42,86 78,95 78,574 47,37 64,29 57,89 85,71 78,95 71,435 52,63 71,43 57,89 78,57 47,37 06 47,37 0 68,42 78,57 78,95 85,717 73,68 42,86 47,37 100 78,95 92,868 68,42 57,14 63,16 78,57 84,21 71,439 68,42 71,43 78,95 71,43 57,89 10010 52,63 64,29 73,68 85,71 100 78,5711 57,89 78,57 89,47 64,29 73,68 92,8612 52,63 35,71 68,42 71,43 84,21 85,7113 68,42 50 57,89 64,29 73,68 71,4314 68,42 78,57 57,89 71,43 73,68 64,2915 57,89 57,14 73,68 64,29 78,95 78,5716 63,16 71,43 47,37 57,14 84,21 85,7117 63,16 85,71 89,47 64,29 89,47 92,8618 73,68 57,14 73,68 64,29 73,68 78,5719 63,16 57,14 68,42 71,43 89,47 57,14

57

Results

20 57,89 50 63,16 50 78,95 78,57Mean accuracy 59,47 58,21 66,05 69,29 77,89 77,14Standard deviation 10,27 18,39 11,22 13,19 10,99 20,53Sensitivity 0.621 - 0.636 - 0.803 -Specificity 0.569 - 0.685 - 0.756 -PPV 0.59 - 0.661 - 0.77 -NPV 0.6 - 0.66 - 0.789 -

7.3.1.7 The mutual information with non-regressed data classifier

Description

The results of the mutual information with non-regressed data classifier are shown in table 7.7 andfigure 7.9. The "best" average classification accuracy is obtained when the features were calculatedusing the third atlas level (234 parcels). The mean accuracy stays stable when more features are used,the highest accuracy is obtained when 10 features are used (mean accuracy = ±49.5%). It should benoted that, as this is a binary classification problem, a random guess between depression or healthycontrol would result in a higher classification accuracy. The compactness of the validation set resultdistributions remains consistent (SD = ±10), the compactness of the optional validation set resultdistributions is very high, especially when all features are used. Only healthy controls are present inthe optional validation set (see section 6.1.2), the disjunction in classification accuracies (both verylow and very high with no average results) shows the fact that the results are defined by the startposition instead of the feature values. No positive outliers are present.

Figure 7.9: Violin plot of the results of the mutual information with non-regressed data classifier.

58

Results

Table 7.7: Best results of the mutual information with non-regressed data classifier (Atlas3).

Iteration Var. features 10 features 20 featuresVal.(%) Opt.(%) Val.(%) Opt.(%) Val.(%) Opt.(%)

1 47,37 50 52,63 42,86 52,63 21,432 57,89 35,71 52,63 21,43 52,63 14,293 57,89 78,57 63,16 21,43 42,11 14,294 52,63 64,29 42,11 42,86 42,11 1005 26,32 42,86 47,37 100 36,84 1006 31,58 28,57 57,89 14,29 57,89 14,297 47,37 50 47,37 100 47,37 14,298 47,37 21,43 36,84 28,57 36,84 1009 57,89 21,43 26,32 71,43 31,58 10010 63,16 42,86 36,84 100 47,37 011 47,37 71,43 36,84 28,57 52,63 7,14312 52,63 64,29 42,11 42,86 42,11 10013 42,11 100 63,16 35,71 31,58 10014 63,16 50 57,89 21,43 42,11 10015 42,11 64,29 57,89 71,43 52,63 016 47,37 7,143 31,58 100 52,63 21,4317 31,58 85,71 36,84 28,57 42,11 21,4318 52,63 64,29 42,11 64,29 52,63 14,2919 57,89 42,86 52,63 64,29 36,84 21,4320 36,84 21,43 31,58 64,29 52,63 42,86Mean accuracy 48,16 50,36 45,79 53,21 45,26 45,36Standard deviation 10,3 23,42 10,8 28,97 7,699 40,99Sensitivity 0.507 - 0.485 - 0.479 -Specificity 0.455 - 0.431 - 0.427 -PPV 0.485 - 0.46 - 0.445 -NPV 0.448 - 0.456 - 0.461 -

7.3.1.8 The correlation with regressed data classifier

Description

The results of the correlation with regressed data classifier are shown in table 7.8 and figure 7.10.The best classification accuracy is obtained when the features were calculated using the third atlaslevel (234 parcels). Both the mean accuracy and the compactness of both the validation and theoptional validation set result distribution increase consistently when more features are used. Thehighest accuracy is reached when all 20 features are used (mean accuracy = ±83%) The compactnesswhen all 20 features are used is extremely high (SD = 6.359 and 8.268 for the validation and optionalvalidation set respectively), reflecting the very high quality and consistency of the feature set. Threepositive outliers are present: the seventh iteration when all features are used (average accuracy =±90.9%, sensitivity = 0.9, specificity = 0.889, PPV = 0.9, NPV = 0.889), the fourteenth iteration when allfeatures are used (average accuracy = ±90.9%, sensitivity = 0.9, specificity = 0.889, PPV = 0.9, NPV =0.889) and the twentieth iteration when 10 features are used (average accuracy = ±88.5%, sensitivity= 0.889, specificity = 0.8, PPV = 0.8, NPV = 0.889). This feature set contains the only iteration wherean accuracy of 100% is reached for the validation set (first iteration).

59

Results

Figure 7.10: Violin plot of the results of the correlation with regressed data classifier.

Table 7.8: Best results of the correlation with regressed data classifier (Atlas3).

Iteration Var. features 10 features 20 featuresVal.(%) Opt.(%) Val.(%) Opt.(%) Val.(%) Opt.(%)

1 68,42 64,29 63,16 71,43 100 78,572 57,89 78,57 63,16 85,71 84,21 78,573 57,89 35,71 68,42 71,43 84,21 78,574 57,89 35,71 84,21 64,29 89,47 85,715 73,68 50 73,68 64,29 78,95 92,866 42,11 50 84,21 85,71 78,95 78,577 73,68 50 63,16 78,57 89,47 92,868 63,16 57,14 84,21 71,43 84,21 92,869 63,16 71,43 57,89 64,29 84,21 71,4310 73,68 35,71 73,68 71,43 73,68 71,4311 42,11 35,71 89,47 71,43 78,95 78,5712 63,16 64,29 73,68 78,57 78,95 85,7113 68,42 35,71 73,68 71,43 84,21 85,7114 42,11 78,57 57,89 71,43 89,47 92,8615 52,63 64,29 84,21 71,43 73,68 92,8616 36,84 57,14 73,68 64,29 84,21 78,5717 47,37 28,57 78,95 85,71 84,21 64,2918 68,42 71,43 73,68 64,29 73,68 85,7119 47,37 78,57 68,42 78,57 78,95 92,8620 68,42 57,14 84,21 92,86 89,47 78,57Mean accuracy 58,42 55 73,68 73,93 83,16 82,86Standard deviation 11,64 15,99 9,267 8,23 6,359 8,268Sensitivity 0.611 - 0.706 - 0.858 -Specificity 0.558 - 0.771 - 0.807 -PPV 0.58 - 0.761 - 0.815 -NPV 0.589 - 0.715 - 0.85 -

60

Results

7.3.1.9 The mutual information with regressed data classifier

Description

The results of the mutual information with regressed data classifier are shown in table 7.9 and fig-ure 7.11. The best classification accuracy is obtained when the features were calculated using thethird atlas level (234 parcels). The mean accuracy as well as the compactness for both the validationan optional validation set result distributions remain fairly similar when more features are used. Thehighest accuracy is obtained when all features are used (mean accuracy =±79%). Drawing conclusionsfrom this fact is di�icult. It is possible (this is not definitive) that the variable amount of featureswas high (between 10 and 20), resulting in similar results. The small increase in mean accuracy andcompactness when all 20 features are used could be explained by the small increase in used features.Three positive outliers are present: the second iteration when 10 features are used (average accuracy=±90.9%), the sixth iteration when a variable amount of features is used (average accuracy =±87.6%,sensitivity = 0.889, specificity = 0.9, PPV = 0.889, NPV = 0.9) and the ninth iteration when a variableamount of features is used (average accuracy = ±87.6%, sensitivity = 0.9, specificity = ). The presenceof two outliers in the variable amount of features column strengthens the suspicion that the variableamount of features that is used was high.

Table 7.9: Best results of the mutual information with regressed data classifier (Atlas3).

Iteration Var. features 10 features 20 featuresVal.(%) Opt.(%) Val.(%) Opt.(%) Val.(%) Opt.(%)

1 78,95 64,29 68,42 78,57 89,47 71,432 78,95 57,14 89,47 92,86 73,68 85,713 63,16 50 73,68 85,71 84,21 64,294 52,63 71,43 73,68 71,43 78,95 78,575 78,95 92,86 78,95 92,86 84,21 92,866 89,47 85,71 78,95 71,43 89,47 64,297 68,42 100 78,95 71,43 84,21 78,578 73,68 64,29 68,42 78,57 89,47 78,579 89,47 85,71 84,21 64,29 68,42 85,7110 73,68 71,43 84,21 71,43 73,68 78,5711 78,95 78,57 84,21 85,71 78,95 92,8612 68,42 50 68,42 78,57 78,95 78,5713 78,95 85,71 68,42 71,43 78,95 85,7114 73,68 85,71 63,16 57,14 78,95 64,2915 73,68 78,57 84,21 71,43 73,68 85,7116 63,16 78,57 63,16 64,29 73,68 78,5717 78,95 71,43 63,16 85,71 78,95 71,4318 84,21 85,71 78,95 64,29 68,42 78,5719 73,68 92,86 73,68 85,71 84,21 78,5720 73,68 71,43 78,95 71,43 73,68 92,86Mean accuracy 74,74 76,07 75,26 75,71 79,21 79,29Standard deviation 8,584 13,6 7,824 9,689 6,332 8,719Sensitivity 0.772 - 0.779 - 0.768 -Specificity 0.724 - 0.727 - 0.817 -PPV 0.735 - 0.74 - 0.805 -NPV 0.761 - 0.767 - 0.78 -

61

Results

Figure 7.11: Violin plot of the results of the mutual information with regressed data classifier.

7.3.2 Combined-feature classifiers

7.3.2.1 The structural feature classifier

Description

The structural feature classifier is obtained by combining all three structural subtype feature sets. Theresults of the structural feature classifier are shown in table 7.10 and figure 7.12. The mean accuracydi�ers significantly between the validation and optional validation set (mean accuracy = 58.68% forthe validation set, mean accuracy = 64.29% for the optional validation set). The compactness is similarfor both validation sets.

Table 7.10: Best results of the structural feature classifier

Iteration 36 featuresVal.(%) Opt.(%)

1 42,11 71,432 47,37 42,863 57,89 71,434 63,16 28,575 52,63 57,146 52,63 64,297 84,21 64,298 57,89 71,439 68,42 71,4310 57,89 57,1411 57,89 78,5712 63,16 42,8613 52,63 71,4314 57,89 50

62

Results

15 78,95 71,4316 63,16 78,5717 42,11 64,2918 52,63 85,7119 57,89 78,5720 63,16 64,29Mean accuracy 58,68 64,29Standard deviation 10,43 14,29Sensitivity 0.613 -Specificity 0.561 -PPV 0.58 -NPV 0.594 -

Figure 7.12: Violin plot of the results of the structural feature classifier.

7.3.2.2 The intensity feature classifier

Description

The intensity feature classifier is obtained by combining the absolute and relative intensity featuresets. The results of the intensity feature classifier are shown in table 7.11 and figure 7.13. Both meanaccuracy and the compactness are similar between the validation and optional validation set. Themean accuracy is ±70%. Two positive outliers are present: the second iteration (average accuracy =±80%, sensitvity = 0.89, specificity = 0.8, ppv = 0.8, npv = 0.89) and the seventeenth iteration (averageaccuracy = ±89.5%, sensitvity = 0.8, specificity = 0.78, ppv = 0.8, npv = 0.78).

63

Results

Table 7.11: Best results of the intensity feature classifier

Iteration 40 featuresVal.(%) Opt.(%)

1 84,21053 64,285712 84,21053 78,571433 68,42105 64,285714 78,94737 505 63,15789 57,142866 78,94737 71,428577 68,42105 85,714298 63,15789 78,571439 73,68421 78,5714310 89,47368 71,4285711 84,21053 64,2857112 63,15789 85,7142913 73,68421 57,1428614 63,15789 64,2857115 63,15789 78,5714316 73,68421 71,4285717 78,94737 10018 57,89474 71,4285719 31,57895 28,5714320 78,94737 71,42857Mean accuracy 71,05263 69,64286Standard deviation 12,94847 14,99642Sensitivity 0.738 -Specificity 0.684 -PPV 0.695 -NPV 0.728 -

Figure 7.13: Violin plot of the results of the intensity feature classifier.

64

Results

7.3.2.3 The connectivity feature classifier

Description

The connectivity feature classifier is obtained by combining the correlation and mutual informationwith regressed data set features. The results of the connectivity feature classifier is shown in table 7.12and figure 7.14. The results of the validation set (both mean accuracy and SD) are be�er comparedto the results of the optional validation set; the increase in standard deviation is the most significant(from 8.95 to 18.38). One positive outlier is present: the fourteenth iteration (average accuracy =±73.5%, sensitivity = 0.7, specificity = 0.67, ppv = 0.7, npv = 0.67).

Figure 7.14: Violin plot of the results of the connectivity feature classifier.

Table 7.12: Best results of the connectivity feature classifier

Iteration 40 featuresVal.(%) Opt.(%)

1 68,42 57,142 84,21 64,293 73,68 64,294 52,63 505 57,89 35,716 68,42 71,437 52,63 508 73,68 85,719 68,42 71,4310 57,89 35,7111 52,63 85,7112 73,68 35,7113 73,68 42,8614 68,42 78,5715 63,16 28,57

65

Results

16 73,68 64,2917 63,16 57,1418 57,89 85,7119 73,68 57,1420 57,89 35,71Mean accuracy 65,79 57,86Standard deviation 8,955 18,38Sensitivity 0.688 -Specificity 0.629 -PPV 0.645 -NPV 0.672 -

7.3.2.4 The intensity and connectivity feature classifier

Description

The intensity and connectivity feature classifier is obtained by combining four feature sets: the ab-solute intensity (atlas3) feature set, the relative intensity (atlas5) feature set, the correlation withregressed data (atlas3) feature set and the mutual information with regressed data (atlas3) feature set.The results of the intensity and connectivity feature classifier is shown in figure 7.13 and figure 7.15.This classifier has the best performance of all trained classifiers. It has a mean accuracy of±88.7% anda mean standard deviation of±6.97. The compactness of both result distributions is very high, showingthe high quality of the feature set. Multiple positive outliers are present: the sixth iteration (averageaccuracy = ±90.2%, sensitivity = 1, specificity = 0.9, ppv = 0.9, nvp = 1), the first, fourth, seventh,eighth, thirteenth, sixteenth and seventeenth iteration (average accuracy =±92.1%, sensitivity = 0.89,specificity = 0.8, ppv = 0.8, npv = 0.89) and the nineteenth iteration (average accuracy = ±94.7%,sensitivity = 0.9, specificity = 0.89, ppv = 0.9, npv = 0.89).

Figure 7.15: Violin plot of the results of the connectivity feature classifier.

66

Results

Table 7.13: Best results of the intensity and connectivity feature classifier.

Iteration 80 featuresVal. Opt.

1 84,21053 1002 78,94737 1003 78,94737 85,714294 84,21053 1005 84,21053 71,428576 94,73684 85,714297 84,21053 1008 84,21053 1009 89,47368 92,8571410 89,47368 78,5714311 89,47368 92,8571412 89,47368 92,8571413 84,21053 10014 89,47368 85,7142915 89,47368 92,8571416 84,21053 10017 84,21053 10018 78,94737 85,7142919 89,47368 10020 73,68421 78,57143Mean accuracy 85,26316 92,14286Standard deviation 5,007648 8,945473Sensitivity 0.879 -Specificity 0.827 -PPV 0.835 -NPV 0.872 -

67

Chapter 8

Discussion

This chapter contains the discussion of the results shown and described in chapter 7. The resultsof each feature subtype will be discussed with respect to other feature subtypes belonging to thesame feature type (intensity, connectivity and structural). Focus will be given to accuracy di�erences,di�erences in compactness of result distributions and changes related to the amount of used features.The results of each feature type will also be discussed with respect to the other feature types, similarfocus as in the previous part is used. The combined feature classifiers are discussed a�erwards. Finallythe influence of the atlas level will be discussed.

8.1 Part 1: Feature type specific

8.1.1 Intensity features

No significant di�erence in accuracy exists between the absolute and relative intensity feature classi-fiers. The accuracy of both classifiers is positively correlated with the amount of features that is usedand an increase of compactness is noticed when all 20 features are used. From a result-based pointof view, no clear preference exists between both feature subtypes (contrary to the clinical relevancepoint of view, see section 5.3.3).

8.1.2 Connectivity features

Two large di�erences are noticeable for the connectivity subtype classifiers: the use of a regressedversus non-regressed data set and the change in accuracy and compactness between correlation-basedand mutual information-based classifiers with respect to the used data set.

The first noticeable di�erence is a significant increase in accuracy for both the correlation-based andmutual information-based classifiers when the regressed data set is used. The increase in accuracyis however much bigger for the mutual information-based classifiers (an increase of ±35% versusan increase of ±6%). A considerable increase in compactness is also noticeable for each featuresubtype classifier when the regressed data set is used. The added value of global signal regression (seesection 5.4.2.1) is clear. The possibility of the introduction of anti-correlated connectivities (see [65])leading to false conclusions is partly negated by the fact that the di�erence between two connectivityvalues is calculated. An increase or decrease of both values would not change the di�erence valuebetween them significantly.

The second noticeable di�erence is the fact that the correlation-based classifiers are able to achieveadequate accuracy results (mean accuracy = ±77.5% when all 20 features are used) when using the

69

Part 2: Feature type comparison

non-regressed data set while the mutual information-based classifiers are not able to do this (meanaccuracy = ±48% when all 20 features are used). The reason why the mutual information feature setis not able to correctly distinguish between the healthy controls and the depression group is counter-intuitive. Generally speaking, high variability of feature values between di�erent people both in thehealthy control group and the depression group is the main reason for the limited achieved accuracies(this is most noticeable in the structural feature sets, see section 8.2). A lack of variation however is themain source of inadequate features when mutual information is used with the non-regressed data set.The range of the group matrices (obtained in the group averaging step in the feature selection process,see section 5.4.2.4) is [4.2542,4.6980] for the healthy controls and [4.2914,4.8101] for the depressiongroup. The di�erence matrix obtained has a range of [-0.2316,0.1158]. This range is too low, whichresults in inadequate features. Mutual information has a possible range of [0,+∞[ and the small rangeof values in the group matrices could indicate the presence of a strong global signal. The high increasein accuracy when the regressed data set is used strengthens this suspicion.

8.1.3 Structural features

Not much is to be said about the comparison of the structural feature subtypes from a result-basedpoint of view. All feature subtypes are inadequate for consistently predicting depression. The reasonfor this is given in section 8.2.

8.2 Part 2: Feature type comparison

The best performing classifiers are trained with the connectivity features (only when the regresseddata set is used, see section 8.1.2). An average accuracy of ±83% is obtained with correlation and±79% is obtained with mutual information; both feature sets have a positive outlier with an accu-racy of ±90.9%. The second best performing classifiers are trained with the intensity features. Anaverage accuracy of ±75% is obtained with absolute intensity and an average accuracy of ±73% isobtained with relative intensity; both feature sets have a positive outlier with an accuracy of ±87%(absolute intensity) and ±88% (relative intensity). The worst performing classifiers are trained withthe structural features. An average accuracy of±59.5% is obtained with le� hemisphere thickness, anaverage accuracy of ±56% is obtained with right hemisphere thickness and an average accuracy of±61% is obtained with parcel volume. No considerable outliers, when compared to the other featuretype classifiers, are present.

A significant di�erence in accuracy is noticeable between the three di�erent feature types. A small(±5% to ±10%) decrease in accuracy is noticed between the connectivity and intensity features. Apossible explanation for this phenomenon could be the fact that both features are calculated from thesame data (fMRI data), but that more information is lost during the feature selection process of theintensity features. Intensity features are calculated by averaging the time series, thus ignoring thevariation of the brain activity through time. Connectivity features do not ignore this information,possibly leading to be�er feature sets. The higher quality of the connectivity feature sets is alsoconfirmed by the higher compactness of the result distributions (SD of ±6.3 versus ±9.8 for thevalidation set and SD of ±8.5 versus ±12 for the optional validation set).

Structural features perform worse than both intensity and connectivity features. Decreases in accu-racy of ±20% to ±30% are present when compared to the two other feature types. Similar decreasesin compactness (SD of ±13 versus ±6.3 or ±9.8 for the validation set and SD of ±20% versus ±8.5or ±12 for the optional validation set) are present. A possible explanation for the large decrease inboth accuracy and compactness is the fact that structural features reflect anatomical changes whileintensity and connectivity reflect functional changes. Changes in cortical thickness or volumes would

70

Discussion

reflect possible atrophy in the brain. Measurable brain atrophy could only be the result of very severeforms of depression where a patient experienced multiple depressive episodes. A simplification hasbeen used in this master’s dissertation: di�erences in severity and duration of depression betweenpatients is ignored. This simplification is needed to reduce the problem to a binary classificationproblem. Statistically significant structural features have been found (see secton 4), showing thepresence of anatomical changes. The rate of change (due to possible atrophy) however is not constantwithin the depression group as some people within the depression group have had more depressiveepisodes/su�er longer from depression than others. If more information about the duration andamount of depressive episodes is known (as well as other factors that could influence anatomicalchanges such as used medications or therapies) structural features could probably be used to be�erpredict both depression and depression severity. This problem could likely be solved using regressionmachine learning techniques.

8.3 Part 3: Combined feature classifiers

Combined feature classifiers are investigated as the inclusion of more features could possibly lead toan increase in accuracy. A first step in the creation of combined feature classifiers is the combinationof the di�erent feature subtypes. Three combined feature classifiers are defined: an intensity featureclassifier, a connectivity feature classifier and a structural feature classifier. The performance ofthese classifiers will be compared to the performance of the best subtype classifiers. All combinedfeature classifiers are trained once with all combined features. Due to the decrease in classificationperformance compared to the subtype feature classifiers, no final combined feature containing alldefined features is trained and tested.

8.3.1 The intensity feature classifier

The performance of the intensity feature classifier is slightly lower than both the absolute and relativeintensity feature classifier (±70% compared to±74.5% and±73% respectively). A possible explanationis the fact that by introducing more features also more noise is introduced. The increased noise makesit more di�icult to correctly distinguish healthy controls and depression patients.

8.3.2 The connectivity feature classifier

The performance of the connectivity feature classifier is much lower than the connectivity subtypeclassifiers (except from the mutual information on the non-regressed data classifier). This is counter-intuitive as the connectivity subtype feature sets are the best performing feature sets, both with veryhigh compactness, and because the intensity feature classifier did not share a similar drop in accuracyand compactness. No clear explanation for this phenomenon can be given.

8.3.3 The structural feature classifier

Contrary to the other combined classifiers, the mean accuracy and compactness is not lower thanthe feature subtype classifiers, but is even increased slightly. The increase however is too small to beconsidered significant.

8.3.4 The intensity and connectivity feature classifier

The intensity and conncectivity feature classifier has the best performance of all classifiers. This isboth intuitive and counter-intuitive. The addition of more features to distinguish di�erent classes

71

Part 4: Performance with respect to atlas level

normally leads to higher classification accuracies, so the aggregation of all high accuracy feature setsinto a single feature set intuitively leads to higher classification results. This is however also counter-intuitive as the other combined feature classifiers built from the feature subtypes that are included inthis feature set (the intensity and connectivity feature classifier, see section 8.3.1 and 8.3.2) have loweraccuracies than the feature subtype classifiers.

This classifier performs very well, but caution should be used when forming conclusions. 80 featuresare present in this feature set and only 73 people are used for training (see section 6.1.3). SVMs, andkernel machines in general, are capable of classification when more features than di�erent traininginstances are present. The use of more features however results in a higher chance of overfi�ing. Toconfirm the ability of this classifier more validation is needed.

8.4 Part 4: Performance with respect to atlas level

A final aspect of the results that is noticed is the fact that four of the six feature subtypes that arecalculated with the Lausanne brain atlas (see section 5.2.1) have the best results when atlas level threeis used (234 parcels). This is shown in figure 8.1 where the mean accuracy of both the validation andoptional validation set (when all 20 features are used) of the absolute and relative intensity as well asthe correlation with regressed data feature set are shown. The average accuracy of each atlas level isdenoted by the thicker brown line. This figure shows that on average the best accuracy is obtainedat atlas level 3 and that the higher atlas levels (atlas level 4 and 5) have a slightly higher accuracythan the lower atles level (atlas 1 and 2). This shows that a parcellation that distinguishes more brainparcels is advantageous for feature calculation.

Atlas level one defines parcels that usually contain multiple unique brain regions; features calculatedfrom these parcels will be the average of several brain regions and therefore will not contain allinformation from these brain regions. Atlas level five defines brain regions that are very small. Slightpositioning variations between di�erent scans due to the preprocessing process have considerableinfluence on the feature values, reducing the consistent di�erences needed for high quality features.The parcels defined by atlas level three circumvent both problems. A possible explanation for thisphenomenon, aside from coincidence, is that the parcel size defined in atlas level three is an optimalmiddle ground. The viability of this hypothesis is strengthened when the absolute intensity featuresand mutual information with regressed data features are discussed (see section 5.3.3.2 and 5.4.3.2).Both feature sets contain several subdivisions of a single brain region (both use atlas level 3), showingthat the subdivision leads to higher classification accuracies, but does not use the highest subdivision(atlas level 5) available as the classification accuracies are again reduced.

Figure 8.1: Violin plot of the results of the connectivity feature classifier.

72

Chapter 9

Conclusion

In this master’s dissertation a computer-aided diagnosis tool/classifier is built that is capable of diag-nosing depression based on an fMRI scan. A data set of 106 people, 60 healthy controls and 46 peoplediagnosed with depression, was used to obtain features and train classifiers.

Three di�erent feature types for classification were explored: intensity features, reflecting the averageactivity in the brain; connectivity features, reflecting the functional connectivity between brain regionsand structural features, reflecting the thickness and volume of the brain. An important principlethat was used in the selection process of the di�erent feature types is the possibility of a simpleinterpretation: each feature can easily be linked to the corresponding brain regions and the clinicalrelevance of each feature can easily be determined by a physician.

Each feature type in itself consisted of several subtypes, a total of nine feature subtypes were definedand used to train the classifier. Two intensity feature subtypes were defined: absolute and relativeintensity features. Two di�erent connectivity measures were used to calculate functional connectivitybetween di�erent brain regions: correlation and mutual information. An extra preprocessing step,global signal regression, was used resulting in two di�erent data sets. Both connectivity measureswere calculated on both data sets, resulting in four di�erent connectivity feature subtypes. Threedi�erent structural feature subtypes were used: le� and right cortical thickness and volume. Supportvector machines were trained multiple times for each feature type resulting in a result distribution foreach feature type and subtype. The properties of the result distributions were analyzed to interpretthe viability of each feature subtype as a distinguishing factor for depression.

Every feature type and subtype did have at least some features that can be linked to depressionand thus have some clinical relevance; some feature subtypes (absolute and relative intensity, le�hemishpere cortical thickness and parcel volume) even have high clinical relevance as most of thefeatures are closely linked with depression. This shows that, even when no prior assumptions aboutthe disease have been made, the found features reflect brain anatomy and activity that are alsofound when no data-driven approach is used. The presence of clinical relevant features increasesthe diagnositc value of the found classifiers.

The best performing feature type is the connectivity feature type. An average accuracy of ±83% isreached (validation set 83.16%, optional validation set 82,86%) when correlation is used on the globalsignal regressed data set, an average accuracy of ±79% is reached (validation set 79.21%, optionalvalidation set 82,86%) when mutual information is used. The highest performing trained classifier withcorrelation that is obtained has an average accuracy of 90.9% (validation set 89.47%, optional validationset 92.86%), this accuracy is reached twice in twenty training cycles. The highest performing trainedclassifier with mutual information that is obtained has an average accuracy of 87.7% (validation set

73

84.21%, optional validation set 92.86%). This shows that functional connectivity changes are highlyreliable features that can be used for a possible computer-aided diagnosis tool for depression.

Intensity features also show some capability of correctly predicting depression with absolute intensityfeature classifiers having an average accuracy of ±75% and relative intensity features having anaverage accuracy of±73%. The best performing classifier has an average accuracy of 87.3% (validationset 89.47%, optional validation set 85.71%) for absolute intensity features.

Di�erences in structural features between healthy controls and patients with depression are provento be statistically significant, but they are not capable of accurately predicting depression. Accuraciesbetween 50% and 60% are obtained using structural features, showing the unreliability of this type offeature for classification.

Combining multiple feature subtypes from a single feature type did not result in an increased accuracy.The final combination of the four best performing feature sets (absolute intensity, relative intensity,correlation with regressed data and mutual information with regressed data) however results in thebest performing classifier. This feature set has an average accuracy of±88.7% and the best performingclassifier has an accuracy of ±94.7% (validation set 89.47%, optional validation set 100%). Cautionshould be used as the classifiers trained with this feature set could have been overfi�ed due to thehigh amount of features.

A final remark that needs to be made is the fact that the size of the data set, while big comparedto similar research [36], is rather small. Further research using other and larger data sets is needed toassess the applicability of the results and conclusions obtained in this master’s dissertation.

Nevertheless it can be concluded that a computer-aided diagnosis tool based on resting state fMRIdata can be a reliant method of diagnosing depression. Tools like this could help mental healthprofessionals with the diagnosis of depression and could even be used in advance of a diagnosticinterview, decreasing the workload of mental health professionals and waiting periods for patients.

Future workThe obtained algorithm for feature selection and the resulting classification systems can be expandedin multiple ways. More feature types, such as frequency and network based features could be in-vestigated. The algorithm could be used to obtain feature sets and classification systems for otherdiseases such as post traumatic stress disorder or schizophrenia. Other data sets could be used tofurther validate and refine the algorithm and classifiers. Task-related fMRI data could be used tobuild a classifier. Other imaging techniques such as EEG or fNIRS could be used together with fMRIto increase the accuracy of the obtained classifiers. Bigger data sets could be used so that morecomplex classification systems, such as random forests or artificial neural networks could be applied.

From a personal point of view the most important expansion would be the inclusion of the ability tonot only diagnose depression, but also correctly predict specific subtypes of depression. The ability todiagnose subtypes of depression that are linked to e�icacy of medication types and treatment optionscould decrease the trail period for the patient. Considerable research needs to be done to make thispossible, but the potential benefit of alternative diagnosis tools in psychology and psychiatry is notto be understated.

74

Bibliography

[1] W. H. Organization et al., “Depression and other common mental disorders: global health esti-mates,” Tech. Rep. (World Health Organization, 2017).

[2] J. P. Hornak, “The basics of mri,” h�p://www. cis. rit. edu/htbooks/mri (2006).

[3] D. W. McRobbie, E. A. Moore, M. J. Graves, and M. R. Prince, MRI from Picture to Proton(Cambridge university press, 2017).

[4] D. P. V. Schuerbeek, “Functional mri lecture 2,” (2018).

[5] D. Saldanha, N. Kumar, V. Ryali, K. Srivastava, and A. Pawar, “Serum serotonin abnormality indepression,” Medical Journal Armed Forces India 65, 108 (2009).

[6] J. Hallas, “Evidence of depression provoked by cardiovascular medication: a prescription se-quence symmetry analysis,” Epidemiology , 478 (1996).

[7] H. Durdle, L. H. Lundahl, C.-E. Johanson, and M. Tancer, “Major depression: the relativecontribution of gender, mdma, and cannabis use,” Depression and anxiety 25, 241 (2008).

[8] C. Heim, D. J. Newport, T. Mletzko, A. H. Miller, and C. B. Nemero�, “The link between childhoodtrauma and depression: insights from hpa axis studies in humans,” Psychoneuroendocrinology 33,693 (2008).

[9] S. R. Penzak, Y. S. Reddy, and S. R. Grimsley, “Depression in patients with hiv infection,” AmericanJournal of Health-System Pharmacy 57, 376 (2000).

[10] J. A. Foster and K.-A. M. Neufeld, “Gut–brain axis: how the microbiome influences anxiety anddepression,” Trends in neurosciences 36, 305 (2013).

[11] K. W. Kelley, J. C. O’Connor, M. A. Lawson, R. Dantzer, S. L. Rodriguez-Zas, and R. H. McCusker,“Aging leads to prolonged duration of inflammation-induced depression-like behavior caused bybacillus calme�e-guerin,” Brain, behavior, and immunity 32, 63 (2013).

[12] L. A. Jelenchick, J. C. Eickho�, and M. A. Moreno, ““facebook depression?” social networking siteuse and depression in older adolescents,” Journal of Adolescent Health 52, 128 (2013).

[13] D. G. Blazer, The Age of Melancholy:" Major Depression" and its Social Origin (Routledge, 2012).

[14] S. Eljamel and K. Slavin, Neurostimulation: Principles and Practice (John Wiley & Sons, 2013).

[15] D. Pagnin, V. de �eiroz, S. Pini, and G. B. Cassano, “E�icacy of ect in depression: a meta-analyticreview,” The journal of ECT 20, 13 (2004).

[16] A. P. Association et al., Diagnostic and statistical manual of mental disorders (DSM-5®) (AmericanPsychiatric Pub, 2013).

75

BIBLIOGRAPHY

[17] D. Draulans, “Vlaamse psychiaters slaan alarm: ’mensen plegen zelfmoord omdat ze niet de juistezorg krijgen’,” Knack (2019).

[18] P. Casteels, “Iedereen in therapie: ’de zoektocht naar een psycholoog zal altijd een lijdenswegzijn.",” Knack , 40–44 (2019).

[19] M. B. First, R. L. Spitzer, M. Gibbon, and J. B. Williams, User’s guide for the Structured clinicalinterview for DSM-IV axis I disorders SCID-I: clinician version (American Psychiatric Pub, 1997).

[20] J. Ventura, R. P. Liberman, M. F. Green, A. Shaner, and J. Mintz, “Training and quality assurancewith the structured clinical interview for dsm-iv (scid-i/p),” Psychiatry research 79, 163 (1998).

[21] K. Wi�kampf, H. van Ravesteijn, K. Baas, H. van de Hoogen, A. Schene, P. Bindels, P. Lucassen,E. van de Lisdonk, and H. van Weert, “The accuracy of patient health questionnaire-9 in detectingdepression and measuring depression severity in high-risk groups in primary care,” Generalhospital psychiatry 31, 451 (2009).

[22] A. Pe�ersson, K. B. Boström, P. Gustavsson, and L. Ekselius, “Which instruments to support di-agnosis of depression have su�icient accuracy? a systematic review,” Nordic journal of psychiatry69, 497 (2015).

[23] B. Burle, L. Spieser, C. Roger, L. Casini, T. Hasbroucq, and F. Vidal, “Spatial and temporalresolutions of eeg: Is it really black and white? a scalp current density view,” International Journalof Psychophysiology 97, 210 (2015).

[24] U. R. Acharya, V. K. Sudarshan, H. Adeli, J. Santhosh, J. E. Koh, S. D. Puthanka�i, and A. Adeli,“A novel depression diagnosis index using nonlinear features in eeg signals,” European neurology74, 79 (2015).

[25] O. Faust, P. C. A. Ang, S. D. Puthanka�il, and P. K. Joseph, “Depression diagnosis support systembased on eeg signal entropies,” Journal of mechanics in medicine and biology 14, 1450035 (2014).

[26] B. Hosseinifard, M. H. Moradi, and R. Rostami, in 2011 19th Iranian Conference on ElectricalEngineering (IEEE, 2011) pp. 1–4.

[27] W. Mumtaz, L. Xia, S. S. A. Ali, M. A. M. Yasin, M. Hussain, and A. S. Malik, “Electroencephalo-gram (eeg)-based computer-aided technique to diagnose major depressive disorder (mdd),”Biomedical Signal Processing and Control 31, 108 (2017).

[28] Y. I. Sheline, D. M. Barch, J. L. Price, M. M. Rundle, S. N. Vaishnavi, A. Z. Snyder, M. A. Mintun,S. Wang, R. S. Coalson, and M. E. Raichle, “The default mode network and self-referentialprocesses in depression,” Proceedings of the National Academy of Sciences 106, 1942 (2009).

[29] J. P. Hamilton, M. Farmer, P. Fogelman, and I. H. Gotlib, “Depressive rumination, the default-mode network, and the dark ma�er of clinical neuroscience,” Biological psychiatry 78, 224 (2015).

[30] X. Wang, Y. Ren, and W. Zhang, “Depression disorder classification of fmri data using sparselow-rank functional brain network and graph-based features,” Computational and mathematicalmethods in medicine 2017 (2017).

[31] R. Bhaumik, L. M. Jenkins, J. R. Gowins, R. H. Jacobs, A. Barba, D. K. Bhaumik, and S. A.Langenecker, “Multivariate pa�ern analysis strategies in detection of remi�ed major depressivedisorder using resting state functional connectivity,” NeuroImage: Clinical 16, 390 (2017).

[32] M. Wei, J. Qin, R. Yan, H. Li, Z. Yao, and Q. Lu, “Identifying major depressive disorder using hurstexponent of resting-state brain networks,” Psychiatry Research: Neuroimaging 214, 306 (2013).

76

BIBLIOGRAPHY

[33] J. R. Sato, J. Moll, S. Green, J. F. Deakin, C. E. Thomaz, and R. Zahn, “Machine learning algorithmaccurately detects fmri signature of vulnerability to major depression,” Psychiatry Research:Neuroimaging 233, 289 (2015).

[34] A. F. Marquand, J. Mourão-Miranda, M. J. Brammer, A. J. Cleare, and C. H. Fu, “Neuroanatomyof verbal working memory as a diagnostic biomarker for depression,” Neuroreport 19, 1507 (2008).

[35] C. H. Fu, J. Mourao-Miranda, S. G. Costafreda, A. Khanna, A. F. Marquand, S. C. Williams,and M. J. Brammer, “Pa�ern classification of sad facial processing: toward the development ofneurobiological markers in depression,” Biological psychiatry 63, 656 (2008).

[36] M. J. Patel, A. Khalaf, and H. J. Aizenstein, “Studying depression using imaging and machinelearning methods,” NeuroImage: Clinical 10, 115 (2016).

[37] W. D. Penny, K. J. Friston, J. T. Ashburner, S. J. Kiebel, and T. E. Nichols, Statistical parametricmapping: the analysis of functional brain images (Elsevier, 2011).

[38] S. Whitfield-Gabrieli and A. Nieto-Castanon, “Conn: a functional connectivity toolbox for corre-lated and anticorrelated brain networks,” Brain connectivity 2, 125 (2012).

[39] S. A. Hue�el, A. W. Song, G. McCarthy, et al., Functional magnetic resonance imaging, Vol. 1(Sinauer Associates Sunderland, MA, 2004).

[40] G. B. Forbes, Human body composition: growth, aging, nutrition, and activity (Springer Science &Business Media, 2012).

[41] W. Greiner, �antum mechanics: an introduction (Springer Science & Business Media, 2011).

[42] R. J. Gillies, I. Robey, and R. A. Gatenby, “Causes and consequences of increased glucosemetabolism of cancers,” Journal of Nuclear Medicine 49, 24S (2008).

[43] G. Bush, J. A. Frazier, S. L. Rauch, L. J. Seidman, P. J. Whalen, M. A. Jenike, B. R. Rosen, andJ. Biederman, “Anterior cingulate cortex dysfunction in a�ention-deficit/hyperactivity disorderrevealed by fmri and the counting stroop,” Biological psychiatry 45, 1542 (1999).

[44] E. Alpaydin, Introduction to machine learning (MIT press, 2009).

[45] R. Duprat, S. Desmyter, K. van Heeringen, D. Van den Abbeele, H. Tandt, J. Bakic, G. Pourtois,J. Dedoncker, M. Vervaet, S. Van Autreve, et al., “Accelerated intermi�ent theta burst stimulationtreatment in medication-resistant major depression: A fast road to remission?” Journal of a�ectivedisorders 200, 6 (2016).

[46] M. Brant-Zawadzki, G. D. Gillan, and W. R. Nitz, “Mp rage: a three-dimensional, t1-weighted,gradient-echo sequence–initial experience in the brain.” Radiology 182, 769 (1992).

[47] J. Ashburner, G. Barnes, C. Chen, J. Daunizeau, G. Flandin, K. Friston, S. Kiebel, J. Kilner, V. Litvak,R. Moran, et al., “Spm12 manual,” URL: h�p://www. fil. ion. ucl. ac. uk/spm/doc/spm12 manual. pdf(2016).

[48] L. Yan, Y. Zhuo, Y. Ye, S. X. Xie, J. An, G. K. Aguirre, and J. Wang, “Physiological origin of low-frequency dri� in blood oxygen level dependent (bold) functional magnetic resonance imaging(fmri),” Magnetic resonance in medicine 61, 819 (2009).

[49] A. C. Evans, D. L. Collins, S. Mills, E. Brown, R. Kelly, and T. M. Peters, in 1993 IEEE conferencerecord nuclear science symposium and medical imaging conference (IEEE, 1993) pp. 1813–1817.

[50] P. Hagmann, L. Cammoun, X. Gigandet, R. Meuli, C. J. Honey, V. J. Wedeen, and O. Sporns,“Mapping the structural core of human cerebral cortex,” PLoS biology 6, e159 (2008).

77

BIBLIOGRAPHY

[51] J. Akiyoshi, K. Hieda, Y. Aoki, and H. Nagayama, “Frontal brain hypoactivity as a biologicalsubstrate of anxiety in patients with panic disorders,” Neuropsychobiology 47, 165 (2003).

[52] D. Hecht, “Depression and the hyperactive right-hemisphere,” Neuroscience research 68, 77 (2010).

[53] J. B. Henriques and R. J. Davidson, “Le� frontal hypoactivation in depression.” Journal of abnormalpsychology 100, 535 (1991).

[54] M. S. George, T. A. Ke�er, and R. M. Post, “Prefrontal cortex dysfunction in clinical depression,”Depression 2, 59 (1994).

[55] W. C. Drevets, J. L. Price, J. R. Simpson Jr, R. D. Todd, T. Reich, M. Vannier, and M. E. Raichle,“Subgenual prefrontal cortex abnormalities in mood disorders,” Nature 386, 824 (1997).

[56] L. R. Baxter, J. M. Schwartz, M. E. Phelps, J. C. Mazzio�a, B. H. Guze, C. E. Selin, R. H. Gerner,and R. M. Sumida, “Reduction of prefrontal cortex glucose metabolism common to three typesof depression,” Archives of general psychiatry 46, 243 (1989).

[57] T. Paus and J. Barre�, “Transcranial magnetic stimulation (tms) of the human frontal cortex:implications for repetitive tms treatment of depression,” Journal of Psychiatry and Neuroscience29, 268 (2004).

[58] Z. Nahas, C. C. Teneback, A. Kozel, A. M. Speer, C. DeBrux, M. Molloy, L. Stallings, K. M. Spicer,G. Arana, D. E. Bohning, et al., “Brain e�ects of tms delivered over prefrontal cortex in depressedadults: role of stimulation frequency and coil–cortex distance,” The Journal of neuropsychiatryand clinical neurosciences 13, 459 (2001).

[59] M. P. Van Den Heuvel and H. E. H. Pol, “Exploring the brain network: a review on resting-statefmri functional connectivity,” European neuropsychopharmacology 20, 519 (2010).

[60] M. G. Kendall, “Rank correlation methods.” (1948).

[61] J. Hlinka, M. Paluš, M. Vejmelka, D. Mantini, and M. Corbe�a, “Functional connectivity inresting-state fmri: is linear correlation su�icient?” Neuroimage 54, 2218 (2011).

[62] T. M. Cover and J. A. Thomas, Elements of information theory (John Wiley & Sons, 2012).

[63] A. E. Desjardins, K. A. Kiehl, and P. F. Liddle, “Removal of confounding e�ects of global signalin functional mri analyses,” Neuroimage 13, 751 (2001).

[64] P. M. Macey, K. E. Macey, R. Kumar, and R. M. Harper, “A method for removal of global e�ectsfrom fmri time series,” Neuroimage 22, 360 (2004).

[65] K. Murphy, R. M. Birn, D. A. Handwerker, T. B. Jones, and P. A. Bande�ini, “The impact ofglobal signal regression on resting state correlations: are anti-correlated networks introduced?”Neuroimage 44, 893 (2009).

[66] Z. S. Saad, S. J. Go�s, K. Murphy, G. Chen, H. J. Jo, A. Martin, and R. W. Cox, “Trouble at rest:how correlation pa�erns and group di�erences become distorted a�er global signal regression,”Brain connectivity 2, 25 (2012).

[67] M. D. Fox, D. Zhang, A. Z. Snyder, and M. E. Raichle, “The global signal and observed anticorre-lated resting state brain networks,” Journal of neurophysiology 101, 3270 (2009).

[68] I. Jolli�e, Principal component analysis (Springer, 2011).

[69] S. G. Disner, C. G. Beevers, E. A. Haigh, and A. T. Beck, “Neural mechanisms of the cognitivemodel of depression,” Nature Reviews Neuroscience 12, 467 (2011).

78

BIBLIOGRAPHY

[70] M. D. Greicius, B. H. Flores, V. Menon, G. H. Glover, H. B. Solvason, H. Kenna, A. L. Reiss, and A. F.Schatzberg, “Resting-state functional connectivity in major depression: abnormally increasedcontributions from subgenual cingulate cortex and thalamus,” Biological psychiatry 62, 429 (2007).

[71] J. P. Ramirez-Mahaluf, J. Perramon, B. Otal, P. Villoslada, and A. Compte, “Subgenual anteriorcingulate cortex controls sadness-induced modulations of cognitive and emotional networkhubs,” Scientific reports 8, 8566 (2018).

[72] A. V. Utevsky, D. V. Smith, and S. A. Hue�el, “Precuneus is a functional core of the default-modenetwork,” Journal of Neuroscience 34, 932 (2014).

[73] N. S. Werner, T. Meindl, J. Materne, R. R. Engel, D. Huber, M. Riedel, M. Reiser, and K. Hennig-Fast, “Functional mri study of memory-related brain regions in patients with depressive disorder,”Journal of a�ective disorders 119, 124 (2009).

[74] E. R. Kenny, J. T. O’brien, D. A. Cousins, J. Richardson, A. J. Thomas, M. J. Firbank, and A. M.Blamire, “Functional connectivity in late-life depression using resting-state functional magneticresonance imaging,” The American Journal of Geriatric Psychiatry 18, 643 (2010).

[75] S. L. Naismith, J. Lagopoulos, P. B. Ward, C. G. Davey, C. Li�le, and I. B. Hickie, “Fronto-striatalcorrelates of impaired implicit sequence learning in major depression: an fmri study,” Journal ofa�ective disorders 125, 256 (2010).

[76] A. Du�a, S. McKie, and J. W. Deakin, “Resting state networks in major depressive disorder,”Psychiatry Research: Neuroimaging 224, 139 (2014).

[77] B. Fischl, “Freesurfer,” Neuroimage 62, 774 (2012).

[78] W. Truong, L. Minuzzi, C. N. Soares, B. N. Frey, A. C. Evans, G. M. Mac�een, and G. B.Hall, “Changes in cortical thickness across the lifespan in major depressive disorder,” PsychiatryResearch: Neuroimaging 214, 204 (2013).

[79] R. S. Mackin, D. Tosun, S. G. Mueller, J.-Y. Lee, P. Insel, N. Schu�, D. Truran-Sacrey, P. Arean, J. C.Nelson, and M. W. Weiner, “Pa�erns of reduced cortical thickness in late-life depression andrelationship to psychotherapeutic response,” The American Journal of Geriatric Psychiatry 21, 794(2013).

[80] M. Niu, Y. Wang, Y. Jia, J. Wang, S. Zhong, J. Lin, Y. Sun, L. Zhao, X. Liu, L. Huang, et al., “Commonand specific abnormalities in cortical thickness in patients with major depressive and bipolardisorders,” EBioMedicine 16, 162 (2017).

[81] P. R. Escalona, B. Early, W. M. McDonald, P. M. Doraiswamy, S. A. Shah, M. M. Husain, O. B.Boyko, G. S. Figiel, E. H. Ellinwood Jr, C. B. Nemero�, et al., “Reduction of cerebellar volume inmajor depression: a controlled mri study,” Depression 1, 156 (1993).

[82] T. Moberget, D. Alnaes, T. Kaufmann, N. T. Doan, A. Cordova-Palomera, L. B. Norbom, J. Rokicki,D. van der Meer, O. A. Andreassen, and L. T. Westlye, “Cerebellar grey ma�er volume is associatedwith cognitive function and psychopathology in adolescence,” Biological psychiatry (2019).

[83] L. Baldaçara, J. G. F. Borgio, A. L. T. d. Lacerda, and A. P. Jackowski, “Cerebellum and psychiatricdisorders,” Brazilian Journal of Psychiatry 30, 281 (2008).

[84] K. R. R. Krishnan, W. M. McDonald, P. R. Escalona, P. M. Doraiswamy, C. Na, M. M. Husain,G. S. Figiel, O. B. Boyko, E. H. Ellinwood, and C. B. Nemero�, “Magnetic resonance imaging ofthe caudate nuclei in depression: preliminary observations,” Archives of general psychiatry 49, 553(1992).

79

BIBLIOGRAPHY

[85] M. J. Kim, J. P. Hamilton, and I. H. Gotlib, “Reduced caudate gray ma�er volume in women withmajor depressive disorder,” Psychiatry Research: Neuroimaging 164, 114 (2008).

[86] D. Robinson, H. Wu, R. A. Munne, M. Ashtari, J. M. J. Alvir, G. Lerner, A. Koreen, K. Cole,and B. Bogerts, “Reduced caudate nucleus volume in obsessive-compulsive disorder,” Archivesof General Psychiatry 52, 393 (1995).

[87] E. Bora, B. Harrison, C. Davey, M. Yücel, and C. Pantelis, “Meta-analysis of volumetric abnor-malities in cortico-striatal-pallidal-thalamic circuits in major depressive disorder,” Psychologicalmedicine 42, 671 (2012).

[88] R. Batuwita and V. Palade, “Class imbalance learning methods for support vector machines,”(2013).

[89] N. Japkowicz, in Proc. of the Int’l Conf. on Artificial Intelligence (2000).

80

Appendices

81

Appendix A

MRI parameters of the fMRI data

82

SIEMENS MAGNETOM TrioTim syngo MR B17

\\USER\Hersenen Research 3T\Psychiatrie\rTMS\RestingStateTA: 10:12 PAT: 3 Voxel size: 3.0×3.0×3.0 mm Rel. SNR: 1.00 SIEMENS: ep2d_pace

PropertiesPrio Recon OffBefore measurementAfter measurementLoad to viewer OnInline movie OffAuto store images OnLoad to stamp segments OffLoad images to graphicsegments

Off

Auto open inline display OffStart measurement withoutfurther preparation

On

Wait for user to start OnStart measurements single

RoutineSlice group 1 Slices 40 Dist. factor 10 % Position R10.6 A3.1 H21.9 Orientation T > C-13.0 > S4.9 Phase enc. dir. A >> P Rotation 0.00 degPhase oversampling 0 %FoV read 192 mmFoV phase 100.0 %Slice thickness 3.0 mmTR 2000 msTE 29.0 msAverages 1Concatenations 1Filter NoneCoil elements HEA;HEP

ContrastMTC OffFlip angle 90 degFat suppr. Fat sat.

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

Averaging mode Long termReconstruction MagnitudeMeasurements 300Delay in TR 0 msMultiple series Off

ResolutionBase resolution 64Phase resolution 100 %Phase partial Fourier OffInterpolation Off

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

PAT mode GRAPPAAccel. factor PE 3Ref. lines PE 36Matrix Coil Mode Auto (Triple)Reference scan mode Separate

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

Distortion Corr. OffPrescan Normalize OffRaw filter OnElliptical filter OffHamming Off

GeometryMulti-slice mode InterleavedSeries Interleaved

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

Special sat. None

SystemBody OffHEP OnHEA On

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

Positioning mode REFTable position HTable position 0 mmMSMA S - C - TSagittal R >> LCoronal A >> PTransversal F >> HCoil Combine Mode Sum of SquaresAutoAlign Head > Brain AtlasAuto Coil Select Default

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

Shim mode StandardAdjust with body coil OffConfirm freq. adjustment OffAssume Silicone Off? Ref. amplitude 1H 0.000 VAdjustment Tolerance AutoAdjust volume Position R10.6 A3.1 H21.9 Orientation T > C-13.0 > S4.9 Rotation 0.00 deg R >> L 192 mm A >> P 192 mm F >> H 132 mm

Physio1st Signal/Mode None

BOLDGLM Statistics OffDynamic t-maps OffStarting ignore meas 0Ignore after transition 0Model transition states OnTemp. highpass filter OnThreshold 4.00Paradigm size 20Meas[1] BaselineMeas[2] BaselineMeas[3] BaselineMeas[4] BaselineMeas[5] BaselineMeas[6] BaselineMeas[7] BaselineMeas[8] BaselineMeas[9] BaselineMeas[10] BaselineMeas[11] ActiveMeas[12] ActiveMeas[13] ActiveMeas[14] ActiveMeas[15] ActiveMeas[16] ActiveMeas[17] ActiveMeas[18] ActiveMeas[19] ActiveMeas[20] ActiveMotion correction OffSpatial filter Off

8/+

SIEMENS MAGNETOM TrioTim syngo MR B17

SequenceIntroduction OnBandwidth 2694 Hz/PxFree echo spacing OffEcho spacing 0.48 ms

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

EPI factor 64RF pulse type NormalGradient mode Fast*

9/+